blurbs.csv

WORDCOUNT ~SOURCELENGTH ~SOURCE ~GOAL ~AGENT ~DATA ~METHODS ~RESULTS ~COMMENTS
67 ~83 ~http://news.mit.edu/2017/computer-system-predicts-products-chemical-reactions-0627 ~Larry Hardesty ~A new computer system that can predict the products of chemical reactions. ~Observations from the 100s of sequences of chemical reactions. ~Applying machine learning algorithms to sequences of chemical reactions in order to yield a better prediction. ~Machine learning approach could aid the design of industrial processes for drug manufacturing. ~This approach of applying machine learning can help greatly in the drug manufacturing industry. 
112 ~135 ~http://www.digitaljournal.com/tech-and-science/technology/microsoft-uses-neural-fuzzing-technique-to-find-software-bugs/article/507593 ~Microsoft. ~Detect errors in programs which could lead to attacks. ~ Previous iterations of "fuzz tests" and "training data". ~"Neural Fuzzing" is used, which utiizes a neural network into fuzz testing. ~The method was successfull which covered more source code and led to less crashes. ~The research is both interesting and promising. My only concern is what if the converse if possible? What if a cyber criminal could use neural networks as a method to "break" these programs? Also, there are elements of software testing which involves interpreting natural language and building test cases, which may be far off yet. At the very least, it demonstrates the potential of neural networks. 
198 ~100 ~https://www.poynter.org/news/how-newsrooms-are-using-machine-learning-make-journalists-lives-easier ~ Researchers at New York Times labs. ~To use machine learning to reduce the tedious jobs that journalists must do when preparing a news story. ~Websites and articles used by journalists. ~The tool, known as Editor, uses tags to identify people, places, concepts, quotes, etc in  a webpage or article so that the factchecking that a journalist must do can be done much  faster and so the journalist's time can be better spent on other stories. The algorithm can learn what story the journalist is researching and so can look more relevant data the more it is used. ~The algorithm is already making journalists jobs much easier, but it struggles with the context of different words. For example, it struggles to identify the contextual difference  of 'Washington' in Denzel Washington and The Washington Post. ~The tool is already working in that aspects of journalist's jobs are being being made less  menial and so they can work on other news stories, thus improving the overall quality of news. Once the tool can be taught contextual differences, it will become even better and could have uses in academia and law as a tool for researching. 
97 ~270 ~http://bigthink.com/design-for-good/this-ai-chatbot-will-get-revenge-on-email-scammers-for-you https://www.netsafe.org.nz/aboutnetsafe/ https://www.theguardian.com/world/2017/nov/10/new-zealand-chatbots-artificial-intelligence-scam-conversations https://www.rescam.org/ ~Netsafe, a non-profit organisation from New Zealand ~Fighting email scams using AI to waste Scammer's time. New Zealand loses 250$NZ annually to cyber crime. ~Data was not disclosed. ~Methods were not disclosed. Re:Scam is said to be trained off scammers emails, making it adapt as scammers change there techniques. ~In first 24 hours, 6000 scam emails were fowarded Re:Scam with 1000 ongoing conversations. ~The bots use humour, typos and slang to make them believable. The more emails RE:scam gets the more it will learn and its vocabulary, intelligence and personality traits will improve.	 
98 ~157 ~Online article on The Economic Times https://economictimes.indiatimes.com/tech/software/soon-eye-movements-can-be-your-new-password/articleshow/16809360.cms ~Oleg Komogortsev ~To improve biometric security by accurately measuring eye movement and tracking. ~Numerous eye tracking data gathered from many subjects in a span of years. ~Algorithms relating to computer vision methodologies, perceptrons, neural networks. ~There have been promising results in this matter. Dr. Komogortsev has made a breakthrough with his research, although there are more to be done to improve biometric security. ~It may be possible to forge biometric signatures, perhaps it is already possible now. Eye tracking has been important for security since this is the future of technology. 
354 ~139 ~By S. Roberts, M. Osbourne, M. Ebden, S. Reece, N. Gibson, S. Aigrain url:http://rsta.royalsocietypublishing.org/content/371/1984/20110550 ~The Royal Society publishing. They take part in publishing papers and educational articles in the areas of mathematical, physical and engineering sciences for educational purposes. ~To introduce the Gaussian processes for time-series data analysis and the conceptual framework of Bayesian modelling in this filed. Bayesian modelling for time-series data is discussed in this paper and the foundations of Bayesian non-parametric modelling are presented for Gaussian processes. ~The authors used knowledge they have learned and developed in the area of Gaussian processess to suggest new equations, formulas and ways of implementing algorithms to further improve the study of machine learning. The time-series analysis was casted into the format of a regression problem, in the form y(x)=f(x)+n, where f() is an unknown function and n is an additive noise process. Covariance algoriths were then carried out to collect data on Gaussian processing data. Graphs were drawn to show the result of the algorithms nad how they had an effect on Gaussian processing information. ~Covariance functions-How they formulate a covariance over arbitary large sets. The domain and knowledge influences design of Gaussian process models are discussed which help provide case exmaples to highlight what should be apporached and what shouldn't in the area of Gaussian process model advancement. ~This paper concludes by presenting a short outline of conceptual nad mathematical basis of GP modelling of time series. The possibilities of further extensions using GPs as cornerstones for further research into this field are discovered to have more complex probalistic models, where the authors noted that more research in this area is possible for numerical intergration, global optimization and much more advancement opportunities in the area of machine learning. ~This paper was interesting as the authors delved into retying to discorver algorithms to further improve covariance functions and how to improve them in the area of machine learning, Should further studies in this continue, Gaussian processess for time-series data analysis could be significantly be improved by further more robust, efficient and faster computing algorithms, to enchance the area of study in machine learning globally. 
141 ~140 ~Fortune.com - Amazon Reportedly Beefing Up AI Capabilities In the Cloud http://fortune.com/2017/11/13/amazon-aws-cloud-ai-machine-learning/ ~Aaron Pressman ~Amazon is looking at upgrade their current Amazon Web Services to make it easier for corporations to conduct big data research on Amazons Cloud. ~Large image-sets. Large audio sets. Corporate documents. Transaction histories. METHOD Amazon will provide their servers with much more capable GPUs in order to speed up the types of Machine Learning operations which many corporations are now interested in such as big data-analysis, audio analysis and image analysis. ~The results from this endeavour will upgrade Amazons Web Services platform to a point where it can compete with more bespoke solutions offered by other companies. ~As expected from a publication such as Fortune, this article presents the Machine Learning field as a new industry with great opportunities for investors willing to buy shares in Amazon. 
212 ~188 ~Example 1: ImageNet Classification with Deep Convolutional Neural Networks http://www.nvidia.com/content/tesla/pdf/machine-learning/imagenet-classification-with-deep-convolutional-nn.pdf  ~  Alex Krizhevsky, Ilya Sutskever, Geoffrey E. Hinton   ~ Train a neural network to classify high resolution images  ~ LSVRC-2010 contest (Contest held to create a image recognition software) 1.2 million labeled high resolution images. ~Using Rectified Linear Units (ReLUs)  along with the overall architecture being constructed of 8 layers with weights, first 5 being conventional layers while the remaining 3 are fully connected. With the sets of 4 pixels of the image being inputted into the system and an output of the top 5 most probably objects to be in the image in ascending order with the most probably at the top. ~  They achieved top-1 error (the top answer error rate) of 37.5 and top 5 error (the fraction of test images for which the correct label is not among the five labels considered most probable by the model) of 17% beating the previous best of 47.1% and 28.2% in the ILSVRC-2010 competition   ~  The results are defiantly impressive and an improvement on the previously impressive results but with a limited dataset although being large at 1.2million it is unsure how it would fare it a more realm world test. 
210 ~162 ~Engadget.com - Counterfeiters are using AI and machine learning to make better fakes https://www.engadget.com/2017/11/10/counterfeit-ai-machine-learning-forgery/ ~Andrew Tarantola ~This article addresses the many ways in which Counterfeiters are using Machine Learning techniques to create a variety of fake media, from ban statements to videos. ~A number of different data-sets were used in this article among them, videos of world leaders talking, videos of sounds being made, and images of counterfeit and real handbags. METHOD To create counterfeit videos face tracking and lip tracking were used to overlay the false lip movements on top of the legitimate video. To create predicted sounds the silent video was compared with a database of videos with sound and the best match was played. The detection of fake handbags simply compared an image set of fake handbags with those taken by the user. A handwriting database was also used to create false signatures. ~All cases were successful, the video produced were very realistic but the recreated face movements and sound. The counterfeit handbag detection was also successful. ~This article points to the very real danger that machine learning techniques can be used to create convincing fake videos and documents. However it also highlights the fact that these same techniques can be used to detect and prevent counterfeiting.  
92 ~102 ~http://www.dailymail.co.uk/sciencetech/article-4928842/AI-computer-transforms-sketches-works-art.html ~Cambridge Consultants ~To transform simple sketches into works of art inspired by Van Gogh, Cezanne and Picasso. Vincent interprets a human drawing and the completes the piece for them. ~Researchers showed the computer thousands of paintings from the Renaissance period to current day. ~Researchers trained the computer on contrast, colour and texture. Researchers used machine learning approaches such as adversarial training, perceptual loss, and end-to-end training of stacked networks. ~Vincent can interpret edges and lines being drawn and pick up where the user left off, producing a complete picture.  ~
76 ~121 ~http://www.tomshardware.com/news/nvidia-invests-deep-instinct-cybersecurity,34992.html and https://www.deepinstinct.com/ ~The DeepInstinct team (Dr. Eli David et al.) ~Improve cybersecurity using deep learning techniques. ~Large amounts of malicious/benign files. ~"Proprietary deep learning technology", or, in the common tongue, "we're not telling you, but it's *even cooler* than regular machine learning." ~They claim the system can defeat 99% of cyberattacks, but it's not clear where that number comes from. ~Since this is all proprietary, not an enormous amount of tangible information was provided, unfortunately. 
75 ~107 ~https://spectrum.ieee.org/tech-talk/robotics/artificial-intelligence/artificial-intelligence-beats-captcha ~Dileep George, cofounder of Vicarious ~To mimic human vision and solve CAPTCHAs, i.e. pass the turing test ~Previously, millions of labeled CAPTCHA images were used as a training set. Vicarious' system required much less training data. Specifically 300 times less ~State-of-the-art deep-learning neural nets were used ~CAPTCHAs are said to be useless if an algorithm can solve it at least 1% of the time. Vicarious' solution had an accuracy rate of 66.6 percent. ~None 
320 ~119 ~https://blog.insightdatascience.com/pitcher-prognosis-using-machine-learning-to-predict-baseball-injuries-7f49b36f88e6 ~No agent listed, study performed by Dr. Carl Wivagg, data scientist at Amazon Alexa. ~Using Machine Learning to Predict Baseball Injuries ~The study chose a classical binary classification format, for each player in each game, the game would be labeled according to whether it preceded an injury for that player ~This study was short on time so the best method was deemed to be random forest, as it doesn’t require labor-intensive feature scalings, it is robust to find outliers, It is sensitive to interactions between variables. the random forest hyperparameters were optimised to maximize the area under an ROC curve, which has two characteristics that make it better than accuracy score for this sort of situation, the value of this metric is still meaningful with greatly imbalanced datasets and there are many more games preceding noninjuries in baseball than games preceding injuries and how a risk-predicting application may be used is not necessarily known before deployment: avoiding false positives may matter more than avoiding false negatives, or vice versa. The area under an ROC curve metric does not require me to know in advance where I will set the threshold for identifying players at risk of injury. ~Four random players were picked and an injury scores was calculated for each game in the season they got injured. All four  displayed high injury scores leading up to the injury; with some showing a very sharp uptick in potential injury just prior to the injury. Some other players had similar but mildly noisier trends, while the remaining player had a consistently high injury score. Whether the final player was that prone to injury permenently or it was a culmination of a lot of games with potential injury was not clear. ~This study was designed to be usable by coaches or athletes as a tool to help identify potential injury patterns, so the results were correlated into an easily usable tool. 
153 ~175 ~The Verge - This interactive map uses machine learning to arrange visually similar fonts https://www.theverge.com/2017/4/27/15454362/ideo-font-map-interactive-typography-tool ~Natt Garun & Kevin Ho ~A machine learning algorithm which matches fonts based on their visual characteristics. ~A large data-set of all freely available fonts provided by Google METHOD This algorithm uses computer vision and ML in order to group fonts into clusters based on their visual styles. It uses aspects such as weight, serif, cursive and others to group. ~The algorithm was very successful, so much so that the creator designed an interactive map of the data which can show users similar fonts to the one they've chosen. ~The article makes a point that the algorithm would be incredibly useful for designs when deciding what type of font they should use for a project. It is a good example of who may use such an algorithm and shows that it was created with a purpose in mind. 
294 ~306 ~CNET: https://www.cnet.com/news/ubereats-knows-what-food-youre-craving-before-you-do/ UberEats Website: https://about.ubereats.com Wiki UberEats: https://en.wikipedia.org/wiki/UberEATS TechRepublic Article for Uber: https://www.techrepublic.com/article/how-data-and-machine-learning-are-part-of-ubers-dna/ ~UberEats, a company that have an online meal ordering and delivery platform launched by San-Francisco, California based "ride-hailing" company Uber Technologies, partnering with restaurants in cities around the world. ~Looks at the users' past selections and choices to make suggestions, akin to how Netflix or any movie streaming service would make movie recommendations.  METHOD No particular method was explicitly mentioned in regards to the way the app make recommendations for the user. According to an web article on TechRepublic, given data about how long its may take to deliver food in different parts of cities at different times of the day. From there, they started building machine learning models that give a more accurate prediction based on the data and not limited computation. ~So if you order one type of food, it will recommend more of that type of food e.g. given in the article, if you tend to order pasta, it will show you "a restaurant's spaghetti, lasagna or tortellini shells". Along with their rating systems, it also will direct the users in selecting restaurants and aid in restaurants learning what dishes resonate with people. It has a 5-star rating system and show feedback over the last 90 days. ~I find it intriguing that Machine Learning can be applicable in areas they may not initially come to mind. The use of Machine Learning used in such services to bring about efficiency in services but also making the customer's experience, in whatever aspect, a lot more personalised. It would be interesting to know what kind of machine learning models were built in order to give a better prediction as to what type of food the user would be interested in. 
212 ~107 ~ https://siliconangle.com/blog/2017/06/18/google-use-ai-machine-learning-tackle-extremist-content-youtube/ ~ YOUTUBE ~ To detect and remove unwanted videos from YouTube ~ All the videos uploaded to YouTube ~ YouTube uses classifier based machine learning algorithms to put it's videos into groups. This algorithm works by analysing the content and comparing it with example which are suitable. An example that this is applied to how YouTube tackles extremist content, where it compares a news report which can contain this type of content and is allowed versus genuine extremist content which get's flagged by the classifer for removing. ~ All videos on youtube are classifed into groups and actions are taken on the videos appropriately. For example some videos are instantly removed, some are punished for breaking policy by having comments turned off or made harder to find and some are not allowed to generate ad revenue. ~ The YouTube algorithm has come under fire in the past year from content creates who are having their videos flagged by the algorithm as unsuitable for advertisers for no good reasons, comments on the videos are starting to effect wether or not a video is advertiser friendly and some very PG content is being confused with more mature content. The algorithm is perhaps too sensitive at the minute and should be retuned somewhat. 
67 ~63 ~https://www.sciencedaily.com/releases/2017/11/171115091819.htm ~University College London ~To show that modelling the human brain via machine learning could help testing drug effects that would not be evident from animal trials ~Large-scale data from stroke patients with indexes for the impact of strokes ~Simulated hypothetical drugs ~The machine learning techniques employed show the full pattern of damage across the brain and allowed theraputic effects to be detected with greater sensitivity ~None 
116 ~99 ~https://www.technologyreview.com/s/545631/how-paypal-boosts-security-with-artificial-intelligence/ ~ PayPal ~To detect fraudsters and money launderers ~Data from the customer's purchasing history, in addition to reviewing patterns of likely fraud. METHODS	 PayPal uses Deep Learning algorithms to analyze patterns to detect fraud. When a pattern is revealed, for example, if sudden strings of many small purchases at convenience stores turn out to be fraud, it's turned into a "feature", or a rule that can be applied in real time to stop purchases that fit this profile. RESULT With each positive detection PayPals fraud detection model becomes more and more proficient. ~This system was first introduced in 2013 and from the 100 first implemented cases of fraud, PayPals model has "learned" over 1000 more!  
84 ~13 ~- Slideshare ~- Quora Data scientist ~- Rank the answers/feeds that they get from the user. Present most interesting stories for a user at a given time Interestingness comprises of topical relevace, social relevance and timeliness. Stories comprises of questions and answers and content cretion prediction. ~- Questions, upvotes,downvotes,topics and answers ~- Methods used in quora are Neural network, Logistic regression and decision tree. ~- Answers were ranked based on their upvotes, social relevance and timliness ~- Still dublicate questions are yet to eradicated. 
269 ~68 ~http://www.information-age.com/machine-learning-big-data-123469622/ ~Banks Machine learning and big data are having a positive impact on customer experience, as well as producing extensive benefits for banks GOALS Combined with machine learning it can be used to identify and profile customers through a host of personal and device characteristics. Furthermore, this can all be done in real time, without any need for conscious input by the end user. Deviations and abnormalities that might indicate a risk can be highlighted and challenged with a far greater degree of speed, subtlety and precision. ~By continuously analysing the vast array of data being generated by digital banking ecosystems, it has become possible to create a unique footprint for every single customer. Furthermore, effective deployment of machine learning and big data can support sophisticated real-time assessment of the risk inherent in every single online transaction. ~By continuously analysing the vast array of data being generated by digital banking ecosystems, it has become possible to create a unique footprint for every single customer. Furthermore, effective deployment of machine learning and big data can support sophisticated real-time assessment of the risk inherent in every single online transaction. ~Analysis of device characteristics is equally sophisticated, including the ability to detect the use of cloaking services to hide an IP address, for example. With all these tools combined it is possible to automatically spot a vast range of anomalous behaviour. Crucially, these capabilities extend far beyond traditional solutions, which are typically based on a relatively limited and inflexible set of fraud indicators. ~Imagining a digital banking experience where we can identify ourselves with absolute certainty, simply by being ourselves, is enticing. 
163 ~130 ~NBC News https://www.nbcnews.com/storyline/the-big-questions/how-machine-learning-revolutionizing-diagnosis-rare-diseases-n700901 ~Face2Gene application ~To identify syndromes and diseases using facial recognition software ~Face2Gene’s system uses a machine-learning algorithm, meaning it learns from every new face it scans. The more data it acquires through its use, scientists hope, the more accurate the diagnoses. ~Face2Gene’s algorithms map points on a patient’s face, compare those points with a database containing points from thousands of other faces, and suggest potential diagnoses. ~The application has given medical professionals an idea of what rare diseases a child may have. With this possible result the medical professional will test for the disease and in many cases it is seen to be true. ~This could be very helpful for madeical professionals but it may be inaccurate, if there isn't enough data to compare to. Face2Gene’s predictions could lead an inexperienced user to false positives; it takes an experienced professional to take the app’s suggestions and interpret them in the context of other clinical symptoms, or to order genetic tests. 
131 ~125 ~https://people.csail.mit.edu/rinard/paper/fse17.genesis.pdf (http://news.mit.edu/2017/bug-repair-system-learns-example-0928) ~Fan Long, Peter Amidon, Martin Rinard ~Process human patches to automatically infer code transforms for automatic patch generation. ~Real-world patches and defects collected from 372 Java project ~Given a set of training pairs representing a program before and after a change the system intends to infer a set of transforms used for the search space. This search space is used when presented with a new program to try to generate a patch based on transforms used to fix bugs found in other projects.  ~The system (Genesis) makes it possible to leverage the combined expertise and patch generation strategies of developers worldwide to automatically patch bugs in new applications. ~The system could prove to be very useful for the immediate detection and repairing of faults in new software projects.  
28 ~112 ~https://www.technologynetworks.com/tn/news/scientists-use-machine-learning-to-analyze-language-in-movies-294179 ~University of Washington ~To use machine learning to quantify how much power and agency a script in a movie gives a character ~800 movie scripts ~None ~None 
22 ~85 ~https://sdtimes.com/microsoft-uses-machine-learning-combat-security-vulnerabilities/ ~Christina Cardoza ~Use AI to find and detect software bugs ~None ~New fuzz testing called neural fuzzing ~Improved code coverage ~None 
247 ~316 ~Engineering Uncertainty Estimation in Neural Networks for Time Series Prediction at Uber By Lingxue Zhu & Nikolay Laptev https://eng.uber.com/neural-networks-uncertainty-estimation/ The article was found through https://medium.mybridge.co/machine-learning-top-10-articles-for-the-past-month-v-oct-2017-c87211085729  ~The agent here is Uber Technologies Inc. They are a global technology company. ~Is to more accuratly and confidently predict Uncertainty Estimation in times series forecasting. The goal is to use the Long Short Term Memory (LSTM) which is a popular time series modeling framework with Bayesian Neural Networks(BNN), to more reduce the uncertainty estimation.  ~Uber has massive amounts of data on their drivers, their users, trips taken etc. As this is a tricky task and alot is dependent on outside variables, they need data on public holidays, weather, city population growth etc. ~They use 3 different methods to calculate prediction uncertainty, the first being Model Uncertainty, followed by Model Mispecification and finally Inherent Noise. It is out of the context of thisreport to delve into the 3 much further.  ~The results were very good as, to quote the article, "Using the MC dropout technique and model misspecification distribution, we developed a simple way to provide uncertainty estimation for a BNN forecast at scale while providing 95 percent uncertainty coverage." ~I can see this being very applicable to a business like Uber, being able to accurately predict how many drivers might be needed on a given night, especially over holidays. With it being at 95% already, it is close enough but no doubt they will continue to improve it.  
190 ~177 ~Infoworld.com How PayPal beats the bad guys with machine learning https://www.infoworld.com/article/2907877/machine-learning/how-paypal-reduces-fraud-with-machine-learning.html ~Paypal ~To detect a fraudulent payments conducted through paypals systems.  ~Paypal has millions of already conducted payments in its archives along with a instances of fraud. It uses this to form a huge test set to train models and conduct supervised learning. ~Paypal combines linear methods (in particular Support Vector Machines), Neural Networks and Deep Learning. It finds that different models are necessary for different situations. It therefore uses all three models and utilises ensemble methods, in particular voting whereby the chosen classification is based on the classification chosen by the majority of the individual models.  ~Paypal manages to detect fraud with a high rate of accuracy. However, this results in a high false positives rate also, meaning that many payments are detected as fraudulent when this isnt the case in reality. Paypal have responded to this by implementing easy ways for users to verify the authenticity of payments (e.g. captchas and security checks) ~Paypal uses similar technology in the credit ratings market in order to determine if an individual is a high risk for a loan.  
76 ~134 ~Huffington Post https://www.huffingtonpost.com/entry/three-real-use-cases-of-machine-learning-in-business_us_593a0e91e4b014ae8c69df37 ~GSK ~They used a natural language and text analytics technology to gain insight into the growing concern of parents to vaccinations ~Parents ~They applied algorithms to find out which fears of parents were related to each other. eg Autism ~They created information for parent's the reduce their fears about vaccinations ~This is very useful as it would promote vaccinations and reduce the number of cases of diseases such as measles, mumps and rubella.  
83 ~173 ~http://www.digitaljournal.com/tech-and-science/science/machine-learning-used-to-assess-chemical-reactions/article/506240 http://pubs.acs.org/doi/10.1021/acs.jpclett.7b02364 ~ A team of researchers from Stony Brook University’s Materials Science and Chemical Engineering Department. ~ To teach a machine learning algorithm to help in increasing the performance of catalysts in order to drive reactions toward desired chemical products more speedily. ~ They used X-ray data that was collected and used by a computer to decipher 3D nanoscale structures. ~ N/A ~ N/A ~ The algorithm results in being able to study the structure/architecture of a catalyst while they are reacting. 
136 ~39 ~https://arxiv.org/pdf/1508.04306v1.pdf ~John R. Hershey, Zhuo Chen, Jonathan Le Roux, Shinji Watanabe ~Develop a system to separate individual sounds out when multiple sources are present. ~Wall Street Journal stories read by people (CSR-I : WSJ0) ~Speech separation was performed by constructing time-domain speech signals based on clusters of time-frequency masks for each speaker. Various clustering methods were used: k-means clustering on each segment; k-means clustering on all segment;  spectral clustering within each segment. ~System outperforms the Oracle NMF algorithm that was used as a baseline for comparisons. Can separate individual voices from a combined source, and while the system was trained on English voices only is able to separate different languages as well.  ~The system could see a lot of use with the increase in home automation/assistant systems such as Google Home and Amazon Echo. 
55 ~108 ~http://www.adweek.com/digital/ad-tech-company-appnexus-just-launched-a-machine-learning-enabled-ad-network/ ~Appnexus an add-tech firm ~Use machine learning to target people better and improve their ad revenue ~No Data Shown ~None Mentioned. ~decrease trade times by 73 percent while receiving 13 percent better performance than when buyers did them manually. ~Changing to try people based targeting like facebook and not just cookie and session based. 
501 ~175 ~Sebastian Raschka, model evaluation, model selection, and algorithm selection in Machine learning https://sebastianraschka.com/blog/2016/model-evaluation-selection-part1.html ~The main goal it to evaluate the performance of the machine learning model. This particular field is important as it helps discover whether or not there is structure in your problem for the algorithms to learn and which algorithms are effective. This reduces the time used in fine tuning the algorithms, in order to make good predictions. In this scenario we run a learning algorithm over a dataset with different settings which will return to us different models and the goal is to pick the best performing models. Due to this reason, the need to estimate the respective performance in order to rank the models against each other is required. ~The data used in this paper can be viewed via this archive. https://archive.ics.uci.edu/ml/datasets/Iris ~There are several different approaches to evaluating models. Cross validation is a method of model validation which is generally better than residual. In cross validation we get an indication of how well the algorithm will work when presented with data it has not seen before. One way of resolving this problem is by splitting the training data into two (training around 2/3 and testing around 1/3). This is known as the holdout method. A function approximator fits a particular function using the training set present only. After this, then its asked to predict the output values in the testing set. The main objective is that it has never seen these values before and so it should give us an unbiased estimate of its performance on new unseen data.  The fraction of the correct predictions constitutes our estimate of the prediction accuracy and because we can see how it generalises to the unseen test data we can conclude that the predictions weren't memorized.  ~The main results are as follows, the iris dataset which containsof 50 Setosa , 50 versicolor  and 550 virginica flowers, the data is divided into  the training set two thirds (which is 100) and the testing set one third (which is 50).  The beneficial advantage of this method is that it better than residual method and no longer to compute. The evaluation may depend highly on which data points end up in the training sets and which end up in the test set. For instance looking at our iris data set assuming that the flowers are distributed uniformly in nature, by splitting the data into training and testing sets we introduced an imbalance in the two datasets. In this case we are running and evaluating the model with imbalanced data, which could cause problems with the data analysis. ~I found it interesting that Machine learning can be used to identify patterns in regards to things in nature, such as that of the flower used in this research paper. This shows how vast and versatile machine learning can be, as well as its wide applicability around various scenarios in the modern, changing and fast IT evolving world of today. 
166 ~95 ~https://www.stevens.edu/news/stevens-makes-major-move-artificial-intelligence-machine-learning ~ Giuseppe Ateniese and a team of researchers at Stevens. ~To probe and ethically hack networks and servers so that they can be improved. ~Networks that use conventional security protocols - millions of these worldwide or locally. ~The team developed a neural network known as GAN - generative adversarial network, a set of algorithms that improve the more times they repeat a task. They used GAN to guess passwords. ~GAN went beyond the successes of the best password guessing tools and managed to crack many supposedly secure passwords. It was able to learn human tendencies and patterns in passwords that we would not even know were there.  ~Because GAN was used in ethical hacking, it can be used to expose our predispositions to certain password patterns, and so can help us to improve our password strengths. However, in the wrong hands, GAN could wreak havoc in the data world as a lot of sensitive and private data could be exposed due to weak passwords. 
182 ~142 ~ https://www.forbes.com/sites/bernardmarr/2016/12/29/4-amazing-ways-facebook-uses-deep-learning-to-learn-everything-about-you/2/#158f79d63090 ~ FACEBOOK ~ Direct users towards products they might want to purchase or pages they may want to follow based on the text they use. ~ All inputted data in Facebook conversations. ~ Facebook use semi-supervised neural network they developed called DeepText. It works by learning to analyze the words users post in the context of the rest of the post and build understandings of how users use words and how the meanings change based on words around them. The neural network learns based on how words are used so it is less effected by variations in spelling or slang. ~ This algorithm works very successfully as you can have a conversation with some people on facebook about a product or service and then later see in a sidebar an advertisement for a similar product or service. ~ Facebook has become a goldmine of user data and reguarly suggest pages and products to users based on conversations they have been having. This can then direct them to websites like amazon which employ their own algorithms come up with more recomendations. 
471 ~175 ~Sebastian Raschka, model evaluation, model selection, and algorithm selection in Machine learning https://sebastianraschka.com/blog/2016/model-evaluation-selection-part1.html ~Sebastian Raschka a machine learning aficianado and a data analyst  GOAL	 Evaluating the performance of machine learning model is important as it helps discover whether or not there is structure in your problem for the algorithms to learn and which algorithms are effective.  This reduces the time used in fine tuning the algorithms, in order to make good predictions.  In this scenario we run a learning algorithm over a dataset with different settings which will return to us different models and the goal is to pick the best performing models. For this reason we need to estimate the respective performance in order to rank the models against each other. ~The iris dataset https://archive.ics.uci.edu/ml/datasets/Iris ~There are different approaches to evaluating models. Cross validation is a method of model validation which is generally better than residual.  In cross validation we get an indication of how well the algorithm will work when presented with data it has not seen before.  One way of solving this problem is to split the training data into two (training 2/3 and testing 1/3). This is called the holdout method.  The function approximator fits a function using the training set only. It is then asked to predict the output values in the testing set.  The idea is that it has never seen these values before and so it should give us an unbiased estimate of its performance on new unseen data.   The fraction of the correct predictions constitutes our estimate of the prediction accuracy and because we can see how it generalises to the unseen test data we can conclude that the predictions were not memorized.  ~Using the Iris dataset which consists of 50 Setosa , 50 versicolor  and 550 virginica flowers, the data is divided into  the training set 2/3 (100) and the testing set 1/3 (50).  The advantage of this method is that it better than residual method and no longer to compute. The evaluation may depend highly on which data points end up in the training sets and which end up in the test set.  For instance looking at our iris data set assuming that the flowers are distributed uniformly in nature, by splitting the data into training and testing sets we introduced an imbalance in the two datasets. In this case we are running and evaluating the model with imbalanced data. ~with many bussinesses turning to machine learnig, its improtant that the predictions made by the algorithms are accurate.  this article is interesing as it points out the importance of having good models to train our algorithms.  both methods inroduced in this article point out important topics that need to be considered when training the algoriths. 
67 ~85 ~https://futurism.com/machine-learning-is-aiding-in-the-fight-against-mental-illness/ ~Carnegie Mellon University and Harvard University ~Detect suicidal ideations ~17 people with suicidal ideations and 17 without. METHOD The subjects were placed inside a MRI machine and monitored how their brains responded when told the words death, cruelty, trouble, carefree, good and praise. RESULT The machine learning algorithm was accurate in 91% of all cases (15/17 of those with suicidal ideations and 16/17 of those without) 
68 ~33 ~ Motic http://www.easyscango.com ~ Motic + Global Good  ~ to use image recognition software and machine to identify and count malaria parasites in a blood smear  ~ they use blood smears to count the number of identifiable parasites linked to malaria ~ not declared  ~ they have built a device that allows for a much faster and definitive detection of septic malaria parasites  ~
232 ~177 ~This article was written by Chuck-Hou Yee and was published by Insight Data Science. https://blog.insightdatascience.com/heart-disease-diagnosis-with-deep-learning-c2d92c27e730 ~This project was undertaken by Chuck-Hou Yee of Imagen Technologies and Insight Data Science. ~To develop a system capable of automatic segmentation of the right ventricle (RV) in images from cardiac magnetic resonance imaging (MRI) datasets. ~The dataset used in this project contains 243 physician-segmented images drawn from the MRIs of 16 patients. The images were 216x256 pixes in size. ~Yee applied affine transformations to the data, such as random rotations, translations, zooms and shears, as well as elastic deformations, which locally stretch and compress the image. This was used to force the network to learn that the RV is a solid, crescent-shaped object that can appear in a variety of orientations. Also, to quantify model performance, Yee used the dice coefficient, which compares a mask X delineating what it thinks is the RV with a mask Y produced by a physician. ~Using these machine learning methods, Yee was able to create models that achieve state of the art segmenting of the right ventricle in cardiac MRIs. ~I feel that the model built here by Yee is important in the advancement of medicine and detecting disease. If it can be implemented into other areas of medicine, it would be of great benefit to the medical community and just shows how important machine learning can be to society. 
213 ~114 ~Example 2 : Mitosis Detection in Breast Cancer Histology Images with Deep Neural Networks  http://people.idsia.ch/~ciresan/data/miccai2013.pdf ~  Dan C. Cires, Alessandro Giusti, Luca M. Gambardella, Jurgen Schmidhuber   ~Using deep max-pooling convolutional neural networks to detect mitosis in breast histology images using the pixels in the image to create a patch centered on the pixel  ~The public MITOS dataset including 50 images corresponding to 50 high-power fields in 5 different biopsy slides stained with Hematosin & Eosin   ~Given an input RGB image I, it tries to find a set D = {d1, d2, . . . , dN } of detections, each reporting the centroid coordinates for a single mitosis. After training the program. Each pixel is assigned one of two possible classes, mitosis or non-mitosis, the former to pixels at (or close to) mitosis centroids, the latter to all other pixels.  ~  Their approach yields an F-score of 0.782, significantly higher than the F-score obtained by the closest competitor (0.718) Also won the ICPR 2012 mitosis detection competition, outperforming other contestants by a significant margin, Sample size was only 50 images, although there are multiple mitosis event in each image it is still relatively small compared to other machine learning applications. 
450 ~110 ~b Shengji://www.nytimes.com/2017/05/23/business/google-deepmind-alphago-go-champion-defeat.html by Paul Mozur ~The agent here is called AphaGo. It was creted by a company called DeepMind that were later acquired by Google's parent company AlphaBet. ~The goal here was to beat China's leading player at the ancient game of Go. It had in the past 18 months beaten South Korea's and at the time the World'ls leading player in the game of Go. Shock waves had been felt in China when AphaGo had beaten Mr Lee of South Korea and now there was a new challenger in Mr Ke of China.  ~The data is the game of Go and what possible moves can be made. AlphaGo is a deep learning Neural Network that has been able to learn the game of Go as it plays. The reason they chose the game Go is because even our best super computers at present could not caclculate all of the moves possible within this game and so a brute force approach is most definitely impossible. The data here is always growing as AlphaGo is always updating its strategy as it learns from each game.  ~The method here is essentially AlphaGo and they way it learns is by using convolutional neural networks. It is considered deep learning because it has many layers, 13 to be exact. It learns in a supervised manner. It relies on 2 different components one being the convolutional networks and the tree search proceedure. "The tree search procedure can be regarded as a brute-force approach, whereas the convolutional networks provide a level on intuition to the game-play." They also sometimes let the system play itself to further improve itself seeing as it has already beaten the best human players.  ~The result has been a great success, as it has inspired many people in China to invest alot more time and money into their own deep learning neural networks. "'AlphaGo truly had a big impact' in China, said Wang Shengjin, a professor at the department of electronic engineering at Tsinghua University in Beijing. "Before, we would be discussing how to apply the technology, but it was hard to be clear exactly how to do it, so AlphaGo gave us a vivid example of that."  ~AlphaGo shook the machine learning World when it beat Mr Lee of South Korea and it has improved itself greatly since then. Adding more vibrations around the World by beating the Chinese champion Mr Ke. It seems to be going from strength to strenght and AlphaGo is only the beginning of the applications for this amazing software, the company DeepMind hopes that it will be ables to assist in curing disesase and illness and many more of humanities toughest challenges.  
213 ~187 ~Article was reported by Will Knight and published by technology review. https://www.technologyreview.com/s/602958/an-ai-ophthalmologist-shows-how-machine-learning-may-transform-medicine/ ~This project taken on by Google researchers. ~To goal was to design and create an AI Ophthalmologist to recognize a common form of eye disease as well as many experts can. ~The algorithm used can look at retinal images and detect diabetic retinopathy as well as a trained ophthalmologist can. The researchers created a training set of 128,000 retinal images classified by at least three ophthalmologists. ~Using the deep learning technique, the researchers developed an algorithm to analyse retinal images. Once the algorithm was trained, using the training set of 128,000 retinal images, the researchers tested its performance on 12,000 images and found that it matched or exceeded the performance of experts in identifying the condition and grading its severity. ~In result, the researchers collaborated with scientists in India, where a clinical trial involving real patients is ongoing. This project involves patients receiving a normal consultation, but their images are also fed into the deep-learning system for comparison. According to Google, results from this trial are not yet ready for publication. ~In my opinion, I believe the use of machine learning in medicine is a great idea, especially for certain areas of the world where the lack of expertise is noticeable. 
132 ~214 ~"Which Bugs Will Hackers Exploit First? Machine Learning Promises a Better Guess" - http://www.defenseone.com/technology/2017/11/which-bugs-will-hackers-exploit-first-machine-learning-promises-better-guess/142621/ ~Researchers from Arizona State University were behind implementing this idea ~In cybersecurity we know most vulnerabilities we just need a better way to know which ones pose an imminent threat. Researchers from Arizona State University have developed a machine-learning model to predict which vulnerabilities are the most likely to cause the next headline-grabbing incident. ~Researchers from Arizona State University say their meth 266% better than the CVSS methodology at predicting whether a bug will be exploited. ~The algorithm they created uses web-crawling algorithms and random forest machine learning to search for discussions of new exploits on the dark web, a portion of the Internet that is only accessible via a special browser like Tor, to protect the anonymi 
187 ~157 ~Wireless Communications and Mobile Computing Conference (IWCMC), 2013 9th International, pg 1666 http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=6583806 ~Brandon Amos, Hamilton Turner, Jules White GAOL Dynamically detect malware in Android Applications ~A critical challenge is the need for the collection and experimentation with a large dataset for training malware classifiers, typically spanning hundreds of applications and thousands of feature vectors. These datasets can be difficult to collect accurately, as there is an inherent tradeoff between profiling malware operating maliciously, such as gaining network access on a mobile device, and ensuring that both the malware remains within its sandbox and the malware profile remains accurate ~They used STREAM to send 10,000 input events to each application in the data set and collect a feature vector every 5 seconds. They collected over thirty features in each vector. Feature vectors collected from the training set of applications were used to create classifiers, and then feature vectors from the testing set are used to evaluate the created malware classifiers. Classification rates from the testing set are based on the 47 testing applications used. ~Results in testing these classifiers show detection rates from 68.75% to 81.25% ~
76 ~139 ~Source:http://news.mit.edu/2017/identifying-optimal-product-prices-0918 AGENT:online retailers GOAL:to increase revenue in online shopping ~Methods:matcheing a product with others with similar charecteristics a relationship between demand and price is predicted with a machine learning algoritm secondly changing the price based on pricing curves for it and then a new curve is applied to help optimize priceing across many products and time periods Results: the methods increased the revenue by about 21 percent in some cases so average revenue was 116% Comments: 
131 ~109 ~https://www.newscientist.com/article/2147472-ai-spots-alzheimers-brain-changes-years-before-symptoms-emerge/ ~Nicola Amoroso,Marianna La Rocca ~To differentiate brains in terms of which one is affected by Alzheimer's. ~The algorithm is trained with 67 MRI scans.38 of them were affected and 29 of them were healthy scans. METHOD They developed an algorithm to notice structural changes in the brain caused by Alzheimer's. ~The algorithm tested with on 148 subjects.Program was able to discriminate healthy brain from a brain that is affected by Alzheimer's with the accuracy of 86 percent. It was also able to tell apart healthy brain from one with MCI(mild cognitive impairment) with 84 percent. ~Alzheimer's disease is still incurable so it is perfectly important to have at least an early diagnose system so the patients and their families can be ready for that and make the individual's life easier. 
189 ~57 ~https://en.wikipedia.org/wiki/Deep_Blue_(chess_computer) ~IBM ~Wanted to create a machine that could play chess better than a chess grandmaster. ~Had the machine play against numerous pro players and recorded the results. ~Had the machine constantly play 1v1 against itself; learning what strategies led to wins more often that not. Once ready, was entered into the World Computer Chess Championship, where it tied for second place. A year later, it was pitted against chess grandmaster Garry Kasparov. ~The machine, named Deep Blue, managed to win the first game of six against Kasparov. However, Kasparov won 3 of the following 5 games, and drew the other two, resulting in a 4-2 win against the machine. A year later, they played again, with Deep Blue winning 3.5-2.5. Deep Blue was then retired. ~This was one of the first big displays of the potential power of machine learning. Chess grandmasters are extremely talented induviduals, capable of planning every move approximately 4 turns ahead. Deep Blue, however, was said to have been capable of planning up to 8 turns ahead, and managed to take the victory. As a fan of chess, I found this story extremely impressive. 
193 ~78 ~https://yourstory.com/2017/11/machine-learning-new-catalyst-higher-education/ ~Higher Education Institutions ~To introduce machine learning into the academic world, allowing a tailored approach for individual students and identifying possible improvements in curricula.  ~Extensive data for each student including assignment submissions, grades, course selection etc and data about each course's curriculum, grading methods and teaching materials.  ~This data can be used for many purposes. One such purpose is identifying historical trends in fail grades in certain courses. The performance of different lecturers teaching the same course could be analysed to identify lecturers who result in high failure rates. Student performance can be analysed on a personal basis in order to provide recommendations on areas in which the student or perhaps even the class as a whole, is struggling. Timetabling issues could also be resolved by employing machine learning to make scheduling decisions and monitor the student quota to room size ratios.  ~ An education system with higher student retention, a more personalised student experience and a sophisticated way of identifying teaching staff who are underperforming in their role. ~This seems like an excellent way to overhaul the university system, which has become more concerned with money than performance. 
145 ~150 ~The Atlantic https://www.theatlantic.com/technology/archive/2014/05/all-the-world-a-track-the-trick-that-makes-googles-self-driving-cars-work/370871/ ~Google  ~To have a working self-driving car ~Pre-loaded route before it sets off, for instance the speed limits of a particular road. ~The car is given the pre-loaded route, in other words, the world when it is empty and the software has to figure out how the world is different from the empty route given. ~This means the software has less to compute and therefore focus on real world objects like other cars. This means it is faster performing actions as it doesn't need to generate the world around it, making the journey smoother. ~Although it seems more cheating as the route is mapped out, the car performs well on routes it knows. If dropped outside the route it may not do as well and the amount of roads to be mapped is very large and may take a long time. 
62 ~97 ~http://research.baidu.com/deep-speech-3:exploring-neural-transducers-end-end-speech-recognition/ ~Baidu ~Baidu is a deep neural network that can generate entirely synthetic human voices that are very difficult to distinguish from genuine human speech.  ~Not Mentioned ~Not Mentioned ~It is able to learn intricate parts of speaking such as cadence, accent, pronunciation and pitch and employs deep learning to allow for uses such as real-time translation and biometric security ~
61 ~141 ~http://www.washington.edu/news/2017/11/13/new-tool-quantifies-power-imbalance-between-female-and-male-characters-in-hollywood-movie-scripts/ ~University of Washington. ~To discern how much power and agency film scripts give to each of their characters. ~Nearly 800 movie scripts. ~Language analysis, specifically on the connotation frames of verbs. ~It was discovered that there is a widespread gender bias in how male and female characters are portrayed. A searchable online database was created showing these biases. ~No comment. 
154 ~96 ~ Department of computer science http://jmlr.org/papers/volume15/srivastava14a/srivastava14a.pdf ~University of Toronto ~Neural networks are a set of algorithms which are set loosely after the human brain. These algorithms are designed to recognise patterns by interpreting data, clustering and labelling inputs. All features in one layer of the network are connected to all features on the next layer. These hidden layers allow them to learn complicated relationships between their input and output. This strategy of learning requires a lot of data for with limited data the network will contain complicated relationship. During the training phase we randomly drop components of neural network (outputs) from a layer of neural networks. The strategy known as dropout is used to minimise over fitting. Over fitting refers to a model that models the training data too well. Over fitting has a negative impact on the models performance on new data. Dropout is a technique to tackle over fitting by randomly drop 
470 ~389 ~Magenta website: https://magenta.tensorflow.org Deep learning Tools: http://www.asimovinstitute.org/analyzing-deep-learning-tools-music/ Machine Learning Paper: Neural Translation by Jointly Learning to align and translate by Dzmitry Bahdanau, Jacob University, Bremen & Kyunghun Cho Youshua Bengio*, University de Montreal (19 May 2016). Link for PDF: https://arxiv.org/pdf/1409.0473.pdf ~Magenta, a Google Brain Project that fuses music and machine learning together. ~The data is from an "NSynth Dataset" that contains musical notes, which can be formatted in JSON or TFRecord files of serialised TensorFlow protocol buffers. The data is broken into 3 parts: Train - a training set of 289,205 examples. Instruments do not overlap with valid or test. Valid - a validation set with 12,678 examples. Instruments do not overlap with train. Test - A test set with 4,096 examples. Instruments do not overlap with train. They have feature encodings, note qualities, instrument classes & families. These come under the 3 pieces of information, source, quality and family. Family of what an instrument belongs to, sonic quality of notes and the source of the sound production of the image which can be acoustic, electronic or synthetic. The NSynth Dataset has 305,979 musical notes, each with a unique pitch, timbre, and envelope, 1,006 instruments from commercial sample libraries. METHOD The methods used here were recurrent neural networks (RNN) and long-term short memory (LSTM) models. The RNN loops connections on the network to hold information across inputs. This is used to learn sequential data like music. The TensorFlow, open-source library software for Machine Intelligence. The recurrent connections in the graph are unrolled into a feed-forward network. The network is trained using a gradient descent technique called back-propagation through time. It predicts the next note given a set of the previous notes. This type is supervised because of the problem of melody generation, and helps predict the next note in a sequence. The way labels can be derived from a dataset of music. In terms of LSTM, you have 2 models, Lookback RNN and Attention RNN. The Lookback RNN has custom inputs and labels. The inputs allow the model to easily recognise patterns that would occur in 1 or 2 bars. Also recognised patterns in relation to where the event takes place and the labels make it easier for the model to repeat rather than have it stored in the RNN cell, which is an LSTM. If wanting to learn longer-term structures, the Attention RNN would be the solution. This model can access previous information without having to store it in the RNN cell, this cell also being an LSTM. This model looks at the last n steps when generating an output for the current step. ~This results in making a program that plays automatically off of the data set it was given with the use of RNN and LSTM. As well as giving a demonstration of neural audio synthesis. ~N/A. 
351 ~368 ~Quartz Media LLC, EGU General Assembly 2017, Applying machine learning to global surface ocean and seabed data to reveal the controls on the distribution of deep-sea sediments.  https://qz.com/577842/scientists-have-used-groundbreaking-technology-to-figure-out-how-the-earth-looked-a-billion-years-ago/ http://meetingorganizer.copernicus.org/EGU2017/EGU2017-10097.pdf ~University of Sydney. Researchers: Adriana Dutkiewicz, Dietmar Muller, University of Sydney, School of Geosciences and Simon OCallaghan, Data 61, CSIRO ~The goal was to make a digital map of the seafloor.  This would help the understanding of how sea-levels and ocean circulation have changed over time, how the oceans are affected by climate change and how, in turn, climate change is influenced by the seafloor. ~The US National Oceanic and Atmospheric Administration had a database of approximately 200,00 seafloor sediment samples that they could use.  The samples had been taken by research cruise ships overe a period of 50 years.  It was discovered though that there was a problem with the database in that many of the research summaries did not match the samples they were attached to.  The team went through the samples and their attached summaries and found 14,500 samples that they could use.  ~The team created a support vector algorithm that used these samples to create a 3D interactive global map of the composition of the seafloor and which also projected the composition of the seafloor in unseen areas.  The team then coupled the data set to submarine topography and grids relating to a number of oceanic factors and applied a probabilistic Gaussian process classifier. They focused on 5 types of sediment/rock and used a five-fold cross-validation, holding back 20% of the date at each iteration, to find out how accurate the machine learning predictions were.   ~An interactive 3D map of the seafloor was created.  They discovered that there were significant differences between the distribution of seafloor sediments and rocks in previous hand drawn maps and those shown in the map.  They also found that the occurrence of the five major types of sediment/rock they had studied could be predicted by sea-surface salinity and sea-surface temperature ~A video of the map they created can be seen on https://youtu.be/mgRHIJjMvMM 
25 ~97 ~https://futurism.com/machine-learning-is-making-it-difficult-to-tell-humans-and-computers-apart/ ~Dom Galeon ~To develop a bot that can solve CAPTCHAs ~None ~Tested bot against CAPTCHAs ~AI can solve CAPTCHA roughly half the time ~None 
191 ~107 ~https://www.huffingtonpost.com/entry/the-future-of-machine-learning-in-finance_us_58d55c99e4b06c3d3d3e6d42 ~ Every major financial institution. ~To use machine learning to optimise profit from trade. ~Past financial data and current news stories related to the stocks. ~The machines use a variety of algorithms to analyse and act on past financial history and current news stories to participate in high-frquency tading and trade in billions of dollars every day. The machines attempt to spot stocks which when bought can be sold quickly for a profit, no matter how small. ~Currently 73 percent of everyday trading is executed by machines, meaning that billions of dollars every single day are entrusted to the decisions that these machines make with regards to buying and selling stocks. ~Clearly the machines are working as financial institutions are using them, but they are  not without their faults. Sometimes they can behave abnormally, such as the 'Flash Crash' in 2010 in which the market fell and then recovered after only 36 minutes due to abnormal behaviour by the machines. The future of machines in stocks is looking to automaticall improve profits and portfolios for companies, eliminating the human risk that is currently part and parcel of the industry. 
82 ~141 ~https://www.forbes.com/sites/julianmitchell/2017/08/22/this-company-uses-ai-to-help-lenders-automate-the-mortgage-loan-process/#470457c02809 ~Unisource ~uses machine learning and artificial intelligence to provide tailored lending solutions for mortgage and real estate agencies ~Uses data mining to get data from courthouse and source records. ~not specified ~Unisource can customize transactional services around the individual needs of each client. Their proprietary technology also allows mortgage and real estate firms to automate the lending process and adapt to regulatory changes while maintaining compliance. ~its a use that after reading about the article i realise how useful this is. 
32 ~85 ~https://sdtimes.com/microsoft-uses-machine-learning-combat-security-vulnerabilities/ ~Microsoft ~Use machine learning to improve software security ~None ~They used a technique called fuzzing to check if a file safe specifically a newly developed kind called neural fuzzing ~Interesting article 
181 ~93 ~Apple Machine Learning Journal https://machinelearning.apple.com/2017/09/12/handwriting.html ~Apple's Handwriting Recognition Team ~Their goal was to examine the recognition of a very large range of Chinese characters inputted by handwriting. They looked at how the addition of extra characters would affect the algorithm's accuracy. These characters would need to be able to be inputted and recognised in real time from an iPhone, iPad, or Apple Watch. ~Set of Hanzi characters in Chinese National Standard GB18030-2005 and standard character set GB2312-80 which only includes of the most used 6,763 characters. ~Input a medium-resolution image of 48x48 pixels representing a handwritten Chinese character. Feed these inputs into a number of feature extraction layers with alternating convolution and subsampling. These inputs are checked against the known data models and a best guess as to identify which character has been hand written. ~The experiments showed that accuracy only degrades slowly as the inventory increases, as long as they use training data of sufficient quality and in sufficient quantity. ~The hand-written inputs currently work very on Apple Devices and this development will aid to making the experience even better. 
254 ~96 ~https://securityboulevard.com/2017/11/how-is-machine-learning-used-in-bitdefender-technologies/ ~Bit Defender, a company that invests some 25 percent of its yearly budget in researching and developing ambitious security project. GOALS Revolutionary ideas that grow into breakthrough technologies are what characterize Bitdefender, a company that invests some 25 percent of its yearly budget in researching and developing ambitious security projects. Bitdefender doesn't fully rely on machine learning technology for detection, instead opting for a layered approach. Machine learning is an indispensable part of our technology security stack, not only by proactively and accurately identifying new and unknown threats, but also by augmenting the detection capabilities of those security technologies. ~Bitdefender has a portfolio of 72 patents in areas such as machine learning, anti-spam, anti-phishing, anti-fraud, antimalware, virtualization, BOX-functionality and hardware design, including 42 delivered in the past three years, and 35 under examination. Ten percent of the patents apply to machine learning in malware detection and online threats, deep learning and anomaly-based detection techniques, strengthening Bitdefender's thought leadership positioning globally. ~Since 2009, the development and training of machine learning algorithms has been a key focus for Bitdefender Laboratories, proving extremely effective in detecting threats in a sophisticated, modern threat landscape. ~The experience of working with machine learning algorithms to detect new and unknown malware samples has substantially improved detection rates and reduced false positives. For Bitdefender, machine learning has proven the best method in data analysis, polymorphic and generic malware detection, among others. ~For Bitdefender, Machine Learning has really paved a way for it to offer the best possible protection to its users. 
242 ~77 ~https://www.engadget.com/2017/11/10/counterfeit-ai-machine-learning-forgery/ ~Authorities who combat counterfeiters and scammers ~To use the same technology which allows counterfeiters, hackers, scammers etc to create fake goods, media and malware in order to identify and stop these counterfeits and fakes.  ~Data on the wide range of counterfeits, scams and fakes on the internet can be analysed and compared to genuines using the same technology creating the counterfeits.  ~For example in the article, an organisation like Entrupy uses machine learning algorithms to identify expensive luxury goods and authenticate them with 98.5% accuracy (approx). These systems are constantly used and constantly gain new data as users try to authenticate them. As someone tries to verify a fake, the service can add the fake to its database and label it as such, for use in future comparisons. The same idea applies when a genuine is identified.  ~By using the same technology which looks at genuines and fakes and tries to make the fake look more like the genuine article, we can distinguish the fake from the genuine. The algorithms used to fake things can be counteractively used to identify the fakes. Such an application is forensics - where forensic analysis has been done solely by hand for a long time. As criminals can now use machine learning to plant evidence, contaminate evidence etc, it is time that law enforcement uses the same technology to identify the planted, contaminated and tampered evidence. ~N/A comments already made through article.  
364 ~104 ~By András Gy"orgy, Tamás Linder, Gábor Lugosi, Gy"orgy Ottucsák url:https://arxiv.org/pdf/0704.1020.pdf ~Cornell University Library. ~The main goal is when given a weighted and directed acyclic graph, where the edge weights can change in a certain arbitrary direction, a decision maker has to choose each turn of a particular game of a certain path to take. The weighted loss of its chosen path in comparison to the weight of its edges are summed up and calculated based on an algorithm for choosing a path effectively. ~Data acquired from the 'multi-armed bandit' case algorithm was studied. This data shows how the decision maker only has access to the loss of the chosen path upon request, where the total number of requests should be bounded by a certain constan 'm'. ~The method the authors used in this paper to evaluate their research is based on a model of the lab-efficient bandit problem for shortest paths, which is motivated by an application to a particular packet switched network model. This model was called the 'cognitive packet network'. These networks contained a type of packet called 'smart packets', which were used to explore the network effectively. These packets would query the delay in networks or loss of data in the network. ~In this paper, the authors have resulted in producing some efficient algorithms for this particular case. These algorithms have been provided for the multi-armed bandit setting and in a combined label efficient multi-armed bandit setting, provided the individual edge losses along the chosen path are revealed to the algorithms. The normalized regrets of the algorithms, compared to the performance of the best fixed path, converge to zero at an O(1/√n) rate as the time horizon n grows to infinity, and increases only polynomially in the number of edges (and vertices) of the graph. ~It is interesting to see how Machine Learning algorithms can be used to further improve and make further advancements into research that solves different cases in the area of shortest path decision making and its choice evaulation. This paper shows how efficient algorithms have been provided for the 'multi-bandit' case and provided methods and algorithms to improve the convergence rate significantly in comparision to other algorithms. 
575 ~119 ~https://arxiv.org/abs/1710.01931 http://yokozunadata.com/ By Anna Guitart, Pei Pei Chen, Paul Bertens, África Periáñez ~Three publicly traded Japanese publishers and a South Korean developer have signed up to use the product. Declining to give their names because of confidentiality agreements.( possibly silicon studios as it is mentioned in the paper) ~The goal of this work is twofold: on the one hand, to accurately forecast time series of in-game sales and playtime; on the other, to simulate events in order to find the best combination of in-game events and the optimal time to publish them.  ~Users gameplay data. such as but not limited to:  Amount of time played What times of day played Game progress Skill level Ages  Sexes ~To achieve the above goals,  we performed an experimental analysis utilizing several techniques such as ARIMA (dynamic regression), gradient boosting, generalized additive models and deep neural networks G. E. Box and G. M. Jenkins, Time series analysis: forecasting and control, revised ed. Holden-Day, 1976 and E. Busseti, I. Osband, and S. Wong, “Deep learning for time series modeling,” Technical report, Stanford University, 2012, J. H. Friedman, “Greedy function approximation: a gradient boosting machine,” Annals of statistics, pp. 1189–1232, 2001, T. J. Hastie and R. J. Tibshirani, Generalized additive models. CRC press, 1990, vol. 43. Pioneering studies on game data science in the field of video games, such as  C. Bauckhage, K. Kersting, R. Sifa, C. Thurau, A. Drachen, and A. Canossa, “How players lose interest in playing a game: An empirical study based on distributions of total playing times,” in Computational Intelligence and Games (CIG), 2012 IEEE conference on. IEEE, 2012, pp. 139–146,  F. Hadiji, R. Sifa, A. Drachen, C. Thurau, K. Kersting, and C. Bauckhage, “Predicting player churn in the wild,” in Computational intelligence and games (CIG), 2014 IEEE conference on. IEEE, 2014, pp. 1–8, A. Peri ´ a´nez, A. Saas, A. Guitart, and C. Magne, “Churn Prediction in ˜ Mobile Social Games: Towards a Complete Assessment Using Survival Ensembles,” in Data Science and Advanced Analytics (DSAA), 2016 IEEE International Conference on. IEEE, 2016, pp. 564–573. P. Bertens, A. Guitart, and A. Peri ´ a´nez, “Games and Big Data: A ˜ Scalable Multi-Dimensional Churn Prediction Model,” Submitted to IEEE CIG, 2017 concentrated in churn prediction. Other related articles that analyze temporal data in the game domain such as: A. Drachen, R. Sifa, C. Bauckhage, and C. Thurau, “Guns, swords and data: Clustering of player behavior in computer games in the wild,” in Computational Intelligence and Games (CIG), 2012 IEEE Conference on. IEEE, 2012, pp. 163–170, A. Drachen, C. Thurau, R. Sifa, and C. Bauckhage, “A comparison of methods for player clustering via behavioral telemetry,” arXiv preprint arXiv:1407.3950, 2014, R. Sifa, C. Bauckhage, and A. Drachen, “The playtime principle: Largescale cross-games interest modeling,” in Computational Intelligence and Games (CIG), 2014 IEEE Conference on. IEEE, 2014, pp. 1–8, A. Saas, A. Guitart, and A. Peri ´ a´nez, “Discovering playing patterns: ˜ Time series clustering of free-to-play game data,” in Computational Intelligence and Games (CIG), 2016 IEEE Conference on. IEEE, 2016, pp. 1–8. focused on unsupervised clustering, not in supervised time series forecast. ~The results suggest that, even though the performance of traditional approaches such as ARIMA is still better, the outcomes of state-of-the-art techniques like deep learning are promising. Deep learning comes up as a well-suited general model that could be used to forecast a variety of time series with different dynamic behaviors. ~
111 ~129 ~https://www.seeker.com/tech/artificial-intelligence/brain-imaging-technology-uses-machine-learning-to-identify-suicidal-thoughts ~Marcel Just, Professor of psychology, Carnegie Mellon University. David Brent, University of Pittsburg. ~To develop brain imaging technology capable of identifying suicidal thoughts. ~Research was carried out on 17 suicidal people and 17 'neurotypical' people, with no history of mental illness, as a control. METHOD: Specifically coded algorithms coupled with an AI system that can detect significant pulses and patterns associated with suicidal thoughts. ~ They were each presented with various keywords relating to death and the brain scanning system distinguished between the two groups with an accuracy rate of 91%. COMMENT: This kind of system could be hugely beneficial for front-line clinicians and could help prevent thousands of suicides. 
89 ~96 ~http://www.newsweek.com/prevent-heart-disease-and-stroke-just-drinking-coffee-709807?yptr=yahoo ~Laura M. Stevens of University of Colorado and other researchers. ~Assessing the health effects of different eating habits, especially heart related health effects. ~They used data from the long-running Framingham Heart Study, which contains information about the participants' eating habits and their health. ~Random decision forests were used to find correlations between different eating habits and health conditions.  ~Drinking coffee was linked to a 7 percent decrease in heart failure and an 8 percent decrease in stroke for every cup drunk per week. ~Correlation doesn't imply causation. 
202 ~100 ~ http://www.techradar.com/news/internet/how-recommendation-algorithms-know-what-you-ll-like-1078924 ~ AMAZON ~ Provide more efficient and accurate customer recomendations ~ Customer history: Viewed, rated and purchased items. ~ Clustering of the customers. The customers are grouped by splitting them into clusters and assign the active customer to a specific cluster. This is done by assigning a vector to each customer. The vector contains data about each item and wether or not the user has interacted with it. The value is positive if the user bought or rated the item, negative if the user disliked the item or zero if the user has not interacted with the item. The clusters were then generated by initially having a random number of empty clusters and assign randomly selected customers to each. Then customers were assigned to clusters based on similarity. Sub algorithms ran during this time to split and merge clusters as they built up. ~ This provided a much more efficient way to quickly generate recomendations for users ~ Here the quality of the recomendations can be low as the purchases and ratings are averaged out in each cluster. Each customer is grouped with a large number of customers which means customers with more unique tastes will often get useless recomendations. 
304 ~120 ~ http://news.psu.edu/story/429727/2016/10/04/research/artificial-intelligence-could-help-farmers-diagnose-crop-diseases ~ Penn State University & The Swiss Federal Institute of Technology(EPFL). Farming & gardening community ~ To bring identification and diagnosis of disease/pest problems in plants to mobile phone devices accessible globally to everyone including those for example in Sub Saharan Countries and other developing countries. Using software similar to Facebook's facial recognition, a leaf is scanned/photographed by the mobile device, processed, and a diagnosis is returned. The classification of these is fast and lightweight, making it possible to bring this ML to mobile phones. The building of the algorithms and the training of the data requires substantial computing power and time.  ~ Publicly accessible library of images of 14 crop species with both healthy and 26 different diseases & symptoms. In total 53,000 images were processed. Images were provided by 'PlantVillage'. ~ A neural network of a large cluster of computers with graphical processing units were used, utilising Deep Learning methods. Each image was categorised into one of 38 classes representing a crop-disease pair. The images were processed so that the network was trained to recognise the patterns in the data associated with the given disease symptoms.  ~ Only 7 out of 1000 images were not correctly classified, in other words a 99.35% success rate was recorded when the PlantVillage data set was used. It was stated that levels of success are likely to be lower in real world situations, but with the growth in mobile phone use internationally, and with the improvements made in their sensors, it is expected that the learning will improve to become highly accurate in time.  ~It's acknowledged that while most farmers can already recognise the problems by eye, this software could help with early detection of a new pest/disease imported from another country, and assist in reducing the likelihood of an outbreak. 
230 ~160 ~https://blog.openai.com/dota-2/, https://blog.openai.com/more-on-dota-2/, and http://www.wildml.com/2017/08/hype-or-not-some-perspective-on-openais-dota-2-bot/ ~The OpenAI research team. ~Defeating professional players in a 1v1 scenario in the game Dota 2. ~The bot had access to the Dota 2 bot API, and hence theoretically had access to lots of information about the game that humans do not have access to. However, besides this and some basic incentives for winning, the bot had no access to any external data; it simply played against itself. ~While we know initial creep blocks were trained using reinforcement learning techniques, they have stated that they're not yet ready to talk about the methods employed. ~The bot sucessully defeated Arteezy (top overall player) and SumaiL (top 1v1) quite decisively, with SumaiL describing the bot as unbeatable. ~There are a few issues with the way the bot is trained: first, the bot only learned to play a single character (there are however over 100 in-game), and secondly, the bot has access to information not available to a human. For example, the bot was able to very accurately determine the range of a particular skill, often using the skill the moment its opponent was in range, and cancelling that skill the moment its opponent was out of range, which is impossible for a human player to replicate. Coupled with the obviously superiour reaction times of the bot, it's not entirely surprising that it would eventually beat a human player. 
52 ~504 ~SOURCEDisease Progression Model(Unsupervised learning) http://people.csail.mit.edu/dsontag/papers/WanSonWan_kdd14.pdfAGENTXiang Wang, David Sontag and Fei WangGOALTo Model disease progression based on real-world evidence .DATA	Medical and health reords of patients.METHODSExpectation Maximization (EM) based algorithm.RESULTSFuture of the medical conditions were predicted with high accuracy.COMMENTSWith larger data sets; increased accuracy of the prediction will aid doctors in saving numerous lives.  
105 ~135 ~https://futurism.com/machine-learning-is-aiding-in-the-fight-against-mental-illness/ https://www.nature.com/articles/s41562-017-0234-y ~Researchers from Carnegie Mellon University and Harvard University. ~Machine learning algorithm trained to understand neural representations of suicidal behaviour. ~17 patients with suicidal ideation and 17 others that were to serve as a control. ~Patients brains were monitored, using a MRI machine, while being presented with 6 key words:  "death, cruelty, trouble, carefree, good, praise" ~Correctly identified 15 out of the 17 patients with suicidal ideation and 16 out of the 17 control Overall accuracy of 91 percent ~While the results of these tests are high, it would be difficult to implement  as it requires the use of an MRI machine. 
83 ~130 ~https://thenextweb.com/artificial-intelligence/2017/11/13/alibabas-ai-is-the-blueprint-for-brick-and-mortar-stores-of-the-future/ ~Alibaba Developers ~To help customers with clothing and accessory suggestions on the items they are trying.Which no real staff can remember every shopping detail of all customers. ~Massive amounts of shopping data ~ML to recognize hundreds of millions of items of clothing with the tastes of designers and fashion experts. ~FashionAI returns dozens of outfit matches for the chosen items of the customer is trying in the same store. ~This is still in trying process but can change the shopping for good. 
191 ~130 ~Quartz article by Jenny Anderson https://qz.com/1094332/mckinsey-used-machine-learning-to-discover-the-best-way-to-teach-science/ ~McKinsey & Company ~The goal of the research discussed was to determine the best of two alternate educational methods. The methods in question are teacher-directed and student-driven. The researchers aimed to provide a scientific answer as to whether or not one is better than the other. ~The data comes from the Organisation for Economic Co-Operation and Development, which tests 15-year-olds around the world on mathematics, reading, and science every three years. The 2015 data set was used in this research. ~The researchers measured the average score increase/decrease of students when utilising the two teaching methods to various degrees, ranging from 'None to Few Lessons' to 'Many to All Lessons' ~It was found that the greatest learning gains occured when many-to-all of the lessons were teacher-directed and some-to-many of the lessons were student-driven. The worst result occurred when none-to-few of the lessons were teacher-directed and many-to-all of the lessons were student-driven. ~There is a caveat in the article, in that the report notes that student-driven teaching increases students' joy in the subject significantly, which is belived to be why a mix of teaching methods is best. 
40 ~122 ~https://www.huffingtonpost.com/entry/the-amazing-ways-how-artificial-intelligence-and-machine_us_59f1b94ee4b09812b938c6e9 ~Stephen Wenge, Carnegie Mellon University ~To predict who is most at risk of a heart attack ~Vast amounts of patient data ~Not Mentioned ~Was able to predict with 80% accuracy whether a patient would have gone into arrest ~N/A 
364 ~348 ~By Trapit Bansal, Igor Mordatch, Jakub Pachocki, Ilya Sutskever & Szymon Sidor. https://arxiv.org/abs/1710.03748 Linked from: https://blog.openai.com/competitive-self-play/?utm_source=mybridge&utm_medium=blog&utm_campaign=read_more Linked from: https://medium.mybridge.co/machine-learning-top-10-articles-for-the-past-month-v-oct-2017-c87211085729 ~Open AI. Which is a non-profit AI research company, discovering and enacting the path to safe artificial general intelligence.  ~To create a enviroment on which an agent can train itself and for the enviroment to change to always allow the  agent to keep evolving beyond basic enviroments. Self-play ensures that the environment is always the right difficulty for an AI to learn. ie, for it not to reach a peak level on an enviroment and stop learning. ~Competitions were set up between multiple simulated 3D robots on a range of basic games. Each agent was trained with  simple goals (push the opponent out of a ring, reach the other side of the ring all while preventing  the other agent from doing the same, to kick the ball into the net or prevent the other agent from doing the same,  and so on). ~PPO - Proximal Policy optimization TRPO - trust region policy optimization The Agents neural network policies were independently trained with PPO. PPO has some of the benefits of TRPO but they are much easier to implement and have better sample complexity (empirically). The experiments test PPO on a collection of benchmark tasks, including simulated robotic locomotion and Atari game playing,  and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable  balance between sample complexity, simplicity, and wall-time. ~By developing agents through thousands of iterations of matches against successively better versions of  themselves, AI systems can be created, that successively bootstrap their own performance; The agent developed new strategies that was not known to them before the start such as  tackling, ducking, faking, kicking and catching, and diving for the ball. ~What was most interesting about this work, was that the agents developed skills unbeknownst to them,  such as tackling, ducking, faking, kicking and catching, and diving for the ball. The process in which they learnt them is obviously trial and error but whats intriquing is they evolved these strategies in the same order that humans would have. 
65 ~97 ~https://qz.com/1094332/mckinsey-used-machine-learning-to-discover-the-best-way-to-teach-science/ ~McKinsey ~To determine the most effective method of teaching; teacher-led or student-driven. ~OECD PISA tests for 2015 which includes more than 500000 students across 72 countries. ~Not discussed ~A mix of the two methods led to better scores with the highest occurring in classes that were teacher-led "most to all"  and student-driven "none-to-some". ~Not conclusive as many question the value of PISA tests. 
72 ~63 ~http://news.mit.edu/2017/miniaturizing-brain-smart-drones-0712 ~Jennifer Chu ~Method for designing efficient computer chips may get miniature smart drones off the ground. DATA	 Performance data of previous chips. ~Using various algorithms for designing efficient computer chips may get miniature smart drones off the ground. ~A computer chip that uses a fraction of power of the larger drone computers and is tailored for a drone as small as a bottlecap. ~This could really help in miniaturising of chips. 
110 ~127 ~HUFFPOST https://www.huffingtonpost.com/entry/three-real-use-cases-of-machine-learning-in-business_us_593a0e91e4b014ae8c69df37 ~WarGaming ~Process large amounts of data received from users and process it to give users real time reults. ~No data mentioned ~Wargaming used Cloudera's advanced analytics platform as a solution for processing over 500 million daily events. ~Using this method allowed Wargaming to enable real-time recommendations to better engage users and present relevant offers. ~Gaming is massive these days and all gaming companies rely on feedback from their users so having a way to process such a large amount of data with the use of machine learning is priceless especially to give quick replies to the users aka gamers who are very impatient in this day and age. 
267 ~35 ~https://eng.uber.com/michelangelo/ ~Uber Engineering develops technologies that create seamless and impactful experiences for our customers. They invest in AI and Machine Learning to complete their vision.  GOALS Uber has contributed to AI and Machine Learning through Michelangelo, a Machine Learning platform that democratizes Machine learning and makes scaling AI to meets the needs of a business easier. Michelangelo makes it easier to enable internal teams to seamlessly build, deploy and operate Machine Learning solutions at Uber's Scale.  ~Michelangelo covers completely the begging to end of Machine Learning: managing data, training, evaluating and deploying models, making predictions and monitoring predictions. The system also supports traditional ML models, time series forecasting and deep learning.  ~Before Michelangelo, Uber faced a number of challenges with building and deploying machine learning modules related to size and scale of operations. Specifically there was no systems in place to build reliable, uniform and reproducible pipelines for creating and managing training and prediction data at scale. Michelangelo address these gaps by standardizing workflows and tools across teams through an end-to-end system that enables users across the company to easily build and operate machine learning systems at scale.  ~Michelangelo began building in mid 2015, Uber started by addressing the challenges around scalable model training and deployment to production serving containers. Then they focused on building better systems for managing and sharing feature pipelines. More recently, the focus shifted to developer productivity - how to speed up the path from idea to first production model and the fast iterations that follow ~I have come to appreciate Michelangelo as a powerful new Machine Learning system.  
123 ~101 ~Google DeepMind https://research.googleblog.com/2016/11/deep-learning-for-detection-of-diabetic.html ~Google DeepMind ~Googles DeepMind project to help doctors to spot the early signs of sight-Threatening eye diseases ~The company's British-based artificial intelligence division will use machine learning to analyse more than one million anonymous eye scans, creating algorithms that can detect early warning signs that humans might miss. The project is DeepMind's second collaboration with the UK's National Health Service (NHS). ~Reinforcement Learning algorithm ~DeepMind's work training an AI agent to teach itself to play Atari games, and thought the company's machine learning skills could be used to analyse eye scans known as Optical Coherence Tomography (OCT) images. ~It is a great step in the history of medical but previously google was accused of accessing patient data without proper authorisation. 
84 ~29 ~https://www.deepinstinct.com ~Deep Instinct CyberSecurity ~Predicts and proactively stops APTs and zero day attacks, as well as more traditional cyber attacks, on any operating system including Android and iOS. ~Deep instinct analyses hundreds of millions of malign files to train its model to detect malware. ~Most malware have the same code except for a few changes from other versions. Typically, these changes are only 2-10% of the code and deep Instinct provides continuous learning allowing for real time cyber security across entire organizations infrastructure. ~~
182 ~214 ~Article by Technology Networks title "Scientists Use Machine-learning to Analyze Language in Movies " https://www.technologynetworks.com/tn/news/scientists-use-machine-learning-to-analyze-language-in-movies-294179 ~Yejin Choi, University of Washington ~Determine levels of agency and power in mover characters based on their script. ~800 movie scripts. ~The study first evaluated the power and agency implicit in 2,000 commonly used verbs. The connotative meanings of these verbs were obtained from Amazon Mechanical Turk crowdsourcing experiments. Power refers to verbs which indicate control over others while agency indicates control over one's own life or storyline.  Specific machine learning technique was not identified with the article simply referring to "machine-learning-based tools" ~Base on the names and descriptions of characters in the scripts they were able to automatically identify genders of 21,000 characters. Male actors spent more time on screen than female actors and also spoke more, accounting for 71.8 percent of the words spoken across all movies. The article reports that there is subtle but widespread gender bias in the way male and female characters are portrayed. Their analysis found that women are consistently portrayed in a way that will reinforce social stereotypes. ~
177 ~111 ~The Guardian https://www.theguardian.com/technology/2016/jul/05/google-deepmind-nhs-machine-learning-blindness ~Google Deep Mind ~Google DeepMind has announced its second collaboration with the NHS, working with Moorfields Eye Hospital in east London to build a machine learning system which will eventually be able to recognise sight-threatening conditions from just a digital scan of the eye. ~At the heart of the research is the sharing of a million anonymous eye scans, which the DeepMind researchers will use to train an algorithm to better spot the early signs of eye conditions such as wet age-related macular degeneration and diabetic retinopathy. ~Training a neural network to do the assessment of eye scans which could vastly increase both the speed and accuracy of diagnosis ~Diabetes can result in a common complication known as diabetic retinopathy that damages the retina. So the goal here is to detect diabetes early with the help of AI to prevent this happening. ~This is the second time deep mind have collaborated with the NHS and critics are concerned about the ways in which patient data would be shared and whether it would remain secure.  
165 ~246 ~Macrumors.com - Deep Neural Networks for Face Detection Explained on Apple's Machine Learning Journal. https://www.macrumors.com/2017/11/16/machine-learning-journal-face-detection/ https://machinelearning.apple.com/2017/11/16/face-detection.html ~Juli Clover & Apple Computer Vision Machine Learning Team. ~This article discusses how Apple uses Machine Learning to provide Face Detection and grouping to users iCloud photos. ~Learned, pre-trained data sets of faces. User photos. METHOD Apple have include face detection in photos since iOS 10 using the Core Image framework. The latest iteration uses Machine Learning to detect faces better in different lighting conditions, as well as group faces and photos by the people in them. Their latest goal is to optimise this procedure for on-device performance by splitting GPU tasks between UI and Deep Neural Network tasks. ~The Apple team have had some success with this method and have managed to reduce the context switching between UI and DNN tasks to less than a millisecond. ~This article offers a much more in-depth analysis of the ML and DNN techniques currently used in industry than many other articles. 
420 ~154 ~Facebook - Deal or no deal? Training AI bots to negotiate https://code.facebook.com/posts/1686672014972296/deal-or-no-deal-training-ai-bots-to-negotiate/ ~Facebook ~To teach chat bots the ability to negotiate i.e. to train a bot to interact and achieve a goal faced with another entity with potentially opposing goals. Specifically, two dialog agents are shown a collection of items, for example 3 bananas, 2 apples and an orange. Each agent has a value function assigned to an item and neither knows the value functions of the other. The aim is for the bots to negotiate so that they both maximise the sum of the values they have assigned to each item.  ~The researchers in this case crowdsourced a collection of negotiations between pairs of people. The negotiation situation replicated that faced by the dialogue bots. The researchers hoped to build dialogue bots that would replicate the behaviour of humans that were successful in their negotiation. ~Dialogue Rollout: At each stage of the conversation, a chat bot would anticipate future resulting conversations that would occur as a result of choosing a certain negotiation tactic. Therefore at each opportunity the bot would take the route which it expected to deliver the highest return later in the negotiation. Recurrent Neural Network: Such a network was used to allow the dialogue bots to be trained to imitate the actions of human negotiators. We have already mentioned how layers in such neural networks maintain a sense of what has happened at previous layers.  Reinforcement Learning: The researchers had the model negotiate with itself also. The model was rewarded when the negotiation goal was achieved, otherwise, the model would be adjusted.  Supervised learning (Neural Networks) were used as a pre-training tactic to map between language and meaning, while reinforcement learning was then utilised in order to achieve the goal of the dialogue bot by choosing between one of the utterances provided by the neural network..  ~The most effective dialogue bot was then tested against humans. Interestingly, most people did not know that they were speaking to a bot when the tests were carried out online. At the same time, the dialogue bot was able to match human negotiators, achieving a better outcome as often as a worse outcome was achieved when compared to a person.  ~An interesting side-effect of the experiment was seeing bots develop their own language to communicate with each other. This raised the concerns of researchers and the public alike of the ability of artificial intelligence to make autonomous decisions, and the potential risks associated to humans.   
150 ~202 ~HealthIT Analytics article titles "Machine Learning 84% Accurate at Flagging Dementia within 2 Years" https://healthitanalytics.com/news/machine-learning-84-accurate-at-flagging-dementia-within-2-years ~A team from the Alzheimer's Disease Neuroimaging Initiative. ~Extend the typical prediction range by several months, reliably predicting cognitive decline up to two years in the future ~A single amyloid positron emission tomography (PET) scan from 273 patients with either age-related stable MCI (sMCI) and those with progressive MCI (pMCI). ~The study used a random forest classifier to classify the risk developing the disease and random undersampling to account for any unevenness in the data set an thus ensuring that both majority and minority classes of patients are represented equally in the training data.  ~In 84% of diagnoses agreed with the human diagnoses ~Undersampling was required as in the methodology as only 15,75% if the 273 patients in the study were diagnosed as having progressive cognitive impairment in the two years following the scan. 
69 ~103 ~https://hackernoon.com/spotifys-discover-weekly-how-machine-learning-finds-your-new-music-19a41ab76efe ~Spotify ~Making personalized 30 song playlist every week called "Discover Weekly" according to user's music taste. ~Spotify’s 140 million users and 30 million songs . ~Collaborative Filtering models,Natural Language Processing models and Audio models. ~"Discover Weekly" is created by fundamentally similar to the songs of your music taste. ~I listen to this playlist and I like almost every song from the songs I hear for the first time. 
598 ~177 ~This article was created by Sophia Ciocca and is available at hackernoon. https://hackernoon.com/spotifys-discover-weekly-how-machine-learning-finds-your-new-music-19a41ab76efe ~The Agent here is "Discover Weekly", a feature from  music streaming company, Spotify that originated in Stockhom, Sweedenand was launched on 7 October 2008. ~The goal here is to create a solid music curation system that recognises each individuals tastes and produces a playlist of 30 songs each Monday for them. There have been music curation attempts made before by other companies such as "Beats", "Songza", "Pandora" and "last.fm" to name afew. Each implementing the own methods with varying degrees of success but none of them ever mastering it. Spotify's goal was to change that by implementing a couple different methods where it's predecessor's only implemented one method.  ~The data they use comes from their own data on their users and their songs. As they use more than one method to acheive their goal they use different types of data for each method. They have 140 million users Worldwide and they have 30 million songs in their database. They also use data from around the web to analyze blogs etc. to see what people are saying about particular artist's or songs. Finally they use indepth data from the audio file. ~The three methods they use are, Collaborative Filtering, Natural Language Processing and Raw Audio Models. With Collaborative Filtering they have a gigantic matrix with each user as a row and each column as a song. It then compares users e.g. user A and user B, and if they have similar tastes it will suggest the songs to user A, that it has not listened to but user B has and vice versa. The NLP method uses words from track metadata, news articles, blogs and other text around the web. Although how exactly they process this data is not known, if it is similar to "Echo Nest", then it groups top terms associated with artists/songs and adds a weight to each term. It then suggests new tracks to users based on the terms. Now comes probably their most important method. Creating Raw Audio Models so comparisons can be done between the audio metadata of the files. This method uses a Convolutional Neural Network similar to the ones that do face recognition although it has been modified to deal with audio data. This method is particularly effective as it means "Discover Weekly" can have songs from artist's that are not so popular. The previous two methods can easily oversee new artist's who have yet to have a surge in popularity and so this is probably the most inmportant method. Some of the characteristics it gets of a song are  key, mode, tempo, loudness and time signature. This can be compared to other tracks user A listens to and if similar can easily end up in user A's "Discover Weekly". ~As a user of Spotify and a big fan of "Discover Weekly", I personally think there results have been a big success. The songs that it puts into "Discover Weekly" are perfect for me and at times I have felt it knows me better than I know myself. The article itself explains how many previous users of Spotify have returned due to how accurate this feature is for and individuals tastes. This to me suggest it is a success and will continue to be as it grows. ~Spotify has many competitors in this space at the moment, like "iTunes" and "Tidal" to name afew. It is features like "Discover Weekly" that set it apart from the rest and keep its fans coming back for more.  
224 ~97 ~https://www.theverge.com/circuitbreaker/2017/11/17/16671328/boston-dynamics-backflip-robot-atlas ~Boston Dynamics GOALS A backflip is a marvel of mechanical engineering and software control. It's a statement of power and poise. ~A humanoid strong enough to jump like that is capable of any "typical" human locomotion. Stairs, curbs, uneven ground, accidental jostling, sitting down, standing up, getting in and out of cars, subway lurches... all moves which are frequently performed by humans who can't land a backflip, and who get mad if you shove them with a hockey stick. ~Performing powerful jumps in a controlled, measured environment is easier than doing dynamic, improvisational parkour. And then humanoids still have to be taught how to do something useful with their newfound physical capabilities. Also, other companies will have to catch up with Boston Dynamics  --  just because this is possible it doesn't mean it's easy.  ~I'm certain there's still much more to do on the software side. Performing powerful jumps in a controlled, measured environment is easier than doing dynamic, improvisational parkour. And then humanoids still have to be taught how to do something useful with their newfound physical capabilities. Also, other companies will have to catch up with Boston Dynamics  --  just because this is possible it doesn't mean it's easy. ~Now we're living in an era where humanoid robots are apparently as agile as we are. Amazing. 
64 ~130 ~Google Research: Classifying facial expressions in VR using eye-tracking cameras (https://research.google.com/pubs/pub46291.html) ~Steven Hickson, Nick Dufour, Avneesh Sud, Vivek Kwatra, Irfan Essa ~Detect facial experession only based on eye-tracking data. ~Image set with: one image with facial expression, two eyes only images. 2000 participant for 50000 images in total. ~CNNs ~A human raters is approximately 50% accurate. The system developed allow around 70% accuracy. ~
66 ~94 ~https://news.stanford.edu/2017/11/15/algorithm-outperforms-radiologists-diagnosing-pneumonia/ ~Stanford Machine Learning Group ~Diagnose diseases based on chest X-ray ~hundreds of thousands of chest X-rays labeled with up to 14 possible pathologies ~Used a deep learning method trained on the labeled x-rays ~At the end it could identify the diseases better than a radiologist in all 14 identification tasks ~The article did not explain the method very well would have liked a better explanation 
234 ~66 ~Fortune http://fortune.com/2015/10/16/how-tesla-autopilot-learns/ ~Tesla ~The goal is that when something unexpected happens, ball rolls onto the street, the car can recognize the pattern and react accordingly (slow down because a child could be running into the street after it). ~Data was gathered from customers driving, data from GPS and maps, and data from company employees driving research cars. 12 sensors on the bottom of the vehicle, a front-facing camera next to the rear-view mirror, and a radar system under the nose. These sensing systems are constantly collecting data to help the autopilot work on the road today, but also to amass data that can make Tesla’s operate better in the future. ~To teach a computer how to take over key parts (or all) of driving using digital sensor systems instead of a human’s senses. To do that companies generally start out by training algorithms using a large amount of data. The algorithms use visual techniques to break down the videos and to understand them. ~Tesla has been able to make high precision maps of the San Francisco bay area from the data gathered and can use these maps to inform its auto pilot system. This allows the car to auto drive on roads that has been mapped. ~From this, Tesla seems to be optimistic and making great strides but is realistic in that there is a long way to go before it is without fault. 
436 ~111 ~http://yokozunadata.com/research/MultiDimChurnPrediction.pdf By Paul Bertens, Anna Guitart and Africa Perianez ~Three publicly traded Japanese publishers and a South Korean developer have signed up to use the product. Declining to give their names because of confidentiality agreements. ~For a comprehensive solution to the churn prediction challenge from several perspectives and dimensions, helping to fully understand and anticipate player attrition. Ie, to predict when a player will leave a game and  how long they will spend playing up to that point. ~The data consisted of player action logs collected between 2014 and 2017 from a major mobile social game, Age of Ishtaria, developed by Silicon Studio.  The predictions were done on a subset of the most valuable players, who provide at least 50% of the revenue (in this case 6.136 players). ~churn prediction in mobile social games A. Perianez, A. Saas, A. Guitart, and C. Magne. Churn prediction in mobile social games: Towards a complete assessment using survival ensembles. In 2016 IEEE International Conference on Data Science and Advanced Analytics (DSAA), pages 564–573, Oct 2016, using conditional inference survival ensembles T. Hothorn, K. Hornik, and A. Zeileis. Unbiased recursive partitioning: A conditional inference framework. Journal of Computational and Graphical Statistics, 15(3):651–674, 2006. Based on survival analysis  T. G. Clark, M. J. Bradburn, S. B. Love, and D. G. Altman. Survival analysis part I: Basic concepts and first analyses. British Journal of Cancer, 89(2):232–238, 2003, the model is capable of performing accurate predictions even when the response variable is censored. Outcome of methods: Two additional models based on A. Perianez, A. Saas, A. Guitart, and C. Magne. Churn prediction in mobile social games: Towards a complete assessment using survival ensembles. In 2016 IEEE International Conference on Data Science and Advanced Analytics (DSAA), pages 564–573, Oct 2016, are implemented to perform predictions on the number of hours a user will play and the level at which they will quit. The models are trained on each of the following response variables: • Playtime: How many seconds the user played the game. • Level: Latest game level reached by the player. In both cases, the censored variable is whether they churned or not (churn is defined as not having logged in for 9 days). ~The results show that the method based on conditional inference survival ensembles is able to model churn both in terms of playtime and level, predicting accurately at which level a player will leave and how long they will play. This indicates that the model is robust to different data distributions, and applicable to different types of response variables ~Won the game data mining competition in 2017. 
106 ~101 ~http://uk.businessinsider.com/visualdx-machine-learning-app-for-skin-diagnosis-ceo-interview-2017-11 ~Dr. Art Papier ~To help diagnose skin conditions and disorders. ~Database of VisualDx which has over 32,000 high-quality medical images some scanned from old collections and films. ~Train the models with VisualDx's own library of professional medical images which is made by expert doctors.It also uses CoreML in iPhones which means photos that are taken with the app for are not stored in the cloud,instead ML algorithm runs on phone to analyse. ~It returns multiple suggestions and a best match with ML. ~VisualDx is intended for use by doctors to confirm and validate diagnoses.It is not yet ready to make a diagnosis all by itself. 
108 ~66 ~https://www.eurekalert.org/pub_releases/2017-10/uoc-mlu102017.php ~ A team of researchers from the University of Cambridge, Los Alamos National Laboratory and Boston University. ~ To develop a method to predict earthquakes. ~ Used simulations of earthquakes to obtain the data needed. ~ They used machine learning to develop a system that studied the relationship between a sound made by a fault line and how close it was to causing an earthquake. ~ The researchers were able to analyse the sound and use it to aid in detecting an earthquake. ~ While these experiments were done using simulations, the research team hopes to scale up the experiments and apply it to real life systems. 
108 ~61 ~http://www.wired.co.uk/article/google-brain-ai-pictures-blur ~Computer Scientists from Google Brain, the centrak Google AI Team ~To add detail and enhance blurry / pixelated images ~Millions of detailed images (of the same resolution) can be used  as a training set, by artifically pixelating them ~The combination of a conditioning neural network and a prior neural network analyzed the images. The conditioning network takes the low-res image and compares it to the high-res images to determing whether a face or a room is in the image ~The AI succesfully added detail and de-blurred images with a small error ~This can be used by intelligence agencies, however they must take into account potential error 
108 ~150 ~Article on Digital Journal by Tim Sandle http://www.digitaljournal.com/business/how-machine-learning-is-shaping-up-investing-interview/article/507704 ~The Machine Learning tool is run by the company Expat Inc. ~The goal of the tool is to predict low, high, opening, and closing prices for stocks in the stock market. ~The tools makes use of market data released by stock exchanges. ~The article does not go into detail, only saying that the tool uses patented machine learning algorithms. ~The tool is able to predict metrics with an accuracy rate between 60 and 80 percent, for the next 30 days. ~All analysis is AI powered and there is no human interference, to remove any element of bias from the analysis. 
40 ~75 ~https://grail.cs.washington.edu/projects/AudioToObama/siggraph17_obama.pdf ~University of Washington ~Synthesise a person's voice and vocal mannerisms ~Footage of President Obama's weekly address ~Association of audio features with mouth shapes, CGI and 3D pose matching ~Photorealistic video of President Obama speaking, generated by the algorithm ~
143 ~286 ~http://www.ibtimes.co.uk/this-neural-network-has-got-some-fascinating-halloween-costume-ideas-1645078 http://www.ibtimes.co.uk/softbank-wants-kick-start-singularity-create-ai-iq-10000-30-years-1644995 http://aiweirdness.com/post/166814009412/a-neural-network-designs-halloween-costumes ~Research scientist Janelle Shane. ~Generating original Halloween costume ideas. ~A dataset was built by crowdsourcing costume ideas from blog readers. ~Neural Networks: Open-source char-rnn written in Torch. Learns each word letter by letter without any extra knowlege its meaning. ~First results weren't great: Sexy sexy Dombie Sexy Cat Sexy A stare Rowan Sexy RoR A the Rog Sexy Cot Sexy Purbie Lampire Poth Rat Reason: Varied dataset so it picked up on most common words first. Eventually it got very good. Punk Tree Disco Monster Spartan Gandalf Starfleet Shark A masked box Martian Devil Panda Clam Potato man Shark Cow Space Batman The shark knight ~Janelle has taken a similar approach to some of her other work. Including a craft beer, a heavy metal band, and my little pony name generators. Examples include: The Fine stranger(beer), Chaosrug(band) and Blue Cuss(pony). 
411 ~210 ~Louis Colombus, Forbes, june,2016 Machine Learning Is Redefining the Enterprise In 2016 https://www.forbes.com/sites/louiscolumbus/2016/06/04/machine-learning-is-redefining-the-enterprise-in-2016/#3f93c43f1871 ~Forbes. Most of the global enterprises have cantered their corporate strategy on digital transformation. Forbes uses applications, algorithms and frameworks to gather information from a huge amount of structured and unstructured data such audio, visuals, texts, voice, body language and facial expression that occur. This developes room for a wider dimension of applications and algorithms ranging from healthcare to video games and even self-driving cars.  ~Using historical data, ML will help businesses grow by basing their decisions on business models with the best outcome.  ML will lead businesses into environmentss with minimum human error and stronger cyber security.ML will completely enable new interactions between customers and companies and eventually allow true intelligent enterprise. ~Historical data and information based on business models, data based on human error, business environments and cyber security. ~Even though machines can automate lengthy and repetitive tasks, thses machines also be used to predict the outcome of new data.  This predictive information data is used to determine and make decisions. Organisations heavily rely on suchpredictions to improve their businesses prospects and reduce the risk taken in decision making as it minimises human error. Behaviour signals in human beings heavily impact prediction algorithms. The more the user shares with the algorithm the better the prediction becomes, one example is streaming platforms such as Netflix and Amazon prime where, depending on the users behaviour, the platform can recommend and advertise content.  These algorithms define a strategy to drive desired sales and market outcomes. Business processes will also become automated and as the business changes these algorithms will constantly update and adapt themselves to the changes in business.  ~Machine Learning is improving enterpirses prospects according to most companies who are applying machine learning to sales, marketing and any relatable area of business in general where machine learning can be applied. With machine learning, companies are gaining a greater predictive accuracy. According to the Accenture institute of higher performance, seventy six percent of the companies that use machine learning claimed to be targeting a higher sales growth in the world of today. ~In the area of business, Machine Learning is highly useful in creating strong predictive algorithms, analysing user preferences, market research opportunies, competitor statistics and other significant areas of improvements. These all contribute to the benefit of improving enterprises in making them more accurate, successful and efficient in the world of business today. 
211 ~112 ~This article was written by Parag Mital and published by Magenta. https://magenta.tensorflow.org/nsynth-fastgen ~This project was undertaken by Parag Mital. ~To implement a fast sampler for NSynth, a neural network audio synthesizer, to make it easier for anyone to generate their own sounds with the model. ~Mital used sounds which he downloaded from freesound.org and then loaded and resampled these to the required sample rate of 16000. ~Mital used audio files which he then encoded and decoded. Once he created the encoding, he then generated(or decoded) it , similar to what an audio player does to an MP3 file. However, since NSynth is such a large model, generating audio can take a few minutes per sample. To solve this, he used deep learning to implement a faster sampler in NSynth. From this, he then explored generating resynthesized, timestretching and interpolation of different audio samples. ~With the help of deep learning, Mital was able to implement a faster sampler in Nsynth, which is now part of Magentas official repo on Github. ~I think the idea of using neural networks which are capable of learning and directly generating raw audio samples, opens up a world of possibilities for music creaters, giving them many more options in their approach to sound design and making music. 
104 ~114 ~http://blog.kaggle.com/2017/05/09/dstl-satellite-imagery-competition-3rd-place-winners-interview-vladimir-sergey/ ~Vladimir Iglovikov, Sergey Mushinskiy ~Detect and label 10 classes of objects including waterways, vehicles, and buildings from satellite imagery. ~Satellite data that is given both in visual and lower frequency regions. Data was divided into train (25 images) and test (32 images) sets. ~Class distributions, Water Classes, Neural Networks, Loss function, Test time augmentation ~We decrease variance of the predictions. Images are split in tiles in a different way and this helps to decrease local boundary effects. ~The Defence and Technology Laboratory challenged people to apply novel techniques to "train the eye in the sky", hence the creation of this competition and results. 
88 ~120 ~Virgin Australia http://www.zdnet.com/article/how-machine-learning-is-helping-virgin-boost-its-frequent-flyer-business/ ~ZDNet ~Help customers to be able to redeem points for experiences. Better predict when is the best time for particular people to redeem points and what should they be redeeming them against. ~Customer data of their preferences and tendencies. ~Build new predictive models at one-tenth of the time it had previously taken. The models are up to 15 percent more accurate than previous ones. ~Virgin can offer more tailored offers and rewards to their customers. ~This is an industrial use of advanced machine learning techniques. 
180 ~105 ~https://www.psychologytoday.com/blog/experimentations/201710/can-artificial-intelligence-predict-suicide ~Marcel Adam Just, Lisa Pan, Vladimir L. Cherkassky, Dana L. McMakin, Christine Cha, Matthew K. Nock & David Brent ~To develop a system that could accurately distinguish patients with suicidal ideations from those without so as  to aid medical professionals in treating them.  ~79 young adults who were either currently experiencing suicidal ideation or were controls with no prior history were assessed for depression, anxiety etc. FMRI was also used during interviews where subjects were asked to actively think about 30 concept 10 related to suicide, 10 positive and 10 negative. ~Data from fMRI was used to identify locations that were similar based on responses to the exercise. This was further distilled into specific locations for neural semantic representation of these concepts which was then fed to their machine learning algorithms which learned how to distinguish between the suicidality, positive and negative emotion represented by a given patients fMRI. ~Model was able to distinguish between those with suicidal ideation and those without and was even able to distinguish between attempters and non-attempters within the ideation group. ~
122 ~79 ~https://blogs.skype.com/news/2014/12/15/skype-translator-how-it-works/?eu=true ~Skype ~To devlop a real-time voice-to-voice translator. ~The training data for speech recognition and machine translation comes from a variety of sources, including translated web pages, videos with captions, as well as previously translated and transcribed one-on-one conversations. ~After the data is prepared and entered into the machine learning system, the machine learning software builds a statistical model of the words in these conversations, and their context. When you say something, the software can find something similar in its statistical model, and apply the previously learned transformation from audio to text and from text into the foreign language ~Skype's voice translator currently works in 8 languages, and their text translator is available in more than 50 languages for instant messaging ~
69 ~35 ~ https://github.com/pakoito/MarI-O ~ SetBLing(youtube) ~ to have an ai use neural networks to learn to pay a level of super mario ~ simulations where ran where the neural network learned to play super mario with no previous knowledge of controls or mechanics  ~ neural networks and genetic algorithms ~ after just 34 attempts Mari/o learns to play the level at the same efficiency of a skilled player ~
138 ~207 ~Example 4: OverFeat: Integrated Recognition, Localization and Detection using Convolutional Networks  http://www.nvidia.com/content/tesla/pdf/machine-learning/overfeat-recognition-localication-detection.pdf ~  Pierre Sermanet David Eigen Xiang Zhang Michael Mathieu Rob Fergus Yann LeCun   ~an integrated framework for using Convolutional Networks for classi- fication, localization and detection. While other attempts have focus on object labeling this attempt was to locate multiple objects and their size in a particular image.  ~Multiple different image collections but mainly referencing using the ImageNet(15million labeled high resolution images) dataset   ~similar to the best ILSVRC12 architecture used by other attempts. But they improved on the network design and the inference steps. By using a separate function to categorize once an object was localized  ~Their entry was the winner of the ILSVRC13 localization competition with 29.9% error (top 5). ~Their result was improved further post competition   
101 ~93 ~http://www.wired.co.uk/article/google-vizier-black-box-optimisation-machine-learning-cookies ~Google Engineers ~To test "Google Vizier" by checking how well it can design cookies ~Baked some recipies in the smaller scale run-through to provide useful data for further baking. ~Automatically optimising hyperparameters of machine learning models. "Transfer learning", effictively learning from experience using data from previous studies. ~While it took a while and few bumps like making the cookies spicy, after a number of rounds the cookies improved significantly and were highly rated and delicious. ~Surprisingly doing something as simply checking if the AI can generate delicious cookies is a great test to see if the AI is well developed. 
53 ~104 ~https://www.forbes.com/sites/quora/2017/04/19/how-does-quora-use-machine-learning-in-2017/#794d6bd83f3a ~Quora ~Its a questions and answers website so they do analysis of the question and answers for things like , understanding,quality,type of question etc. ~The users of the website would provide a lot of the data. ~Logistice Regression,Elastic Nets,(Deep) Neural Networks,k-means and other clustering approaches , Random Forests ~The Quora Platform ~
140 ~81 ~http://www.ibmbigdatahub.com/blog/cyber-security-powered-ai-and-machine-learning ~ IBM ~To develop a Cyber Security Model capable of staying ahead of advances of hackers and cyber criminal activities. ~A "Cyber Security Data Lake" which will augment existing security analytics and anomaly detection solutions whilst incorporating this with additional data sets that are valuable for security intelligence. ~IBM have suggested using Apache Spark, which provides strong framework that can perform batch processing to build a machine learning model from scratch or leveraging existing models from Github. It then uses Spark streaming functionality to apply the intelligence in real-time.  RESULT Enterprises security teams can efficiently build and deploy machine learning models to unstructured and structured data to focus on the discovery of unknown attack vectors. It enables security analysts to deliver insights and data points needed to build the signatures of abnormal behavior beyond traditional security tools. ~NULL 
101 ~81 ~Kyushu University https://boingboing.net/2017/10/31/classifiers-are-fragile.html ~Boing Boing ~Examine how easy it would be to fool a Machine Learning based Artificial Intelligence classification system. ~Dataset of test images. ~Modifying pixels within the test images. ~With only 1 pixel modification, there are 73.8% of the images can be perturbed to one or more target classes, 82.0% and 87.3% in the cases of 3 and 5-pixel attacks. Non-sensitive images are even much rarer than sensitive images even if limiting the perturbation to such a small scope. ~Few-pixel modification is an effective method of searching adversarial images while can be hardly recognised by human eyes in practice. 
51 ~159 ~SOURCE:http://news.mit.edu/2017/faster-big-data-analysis-tensor-algebra-1031 AGENT:Larry Hardesty  GOAL:To achieve faster big-data analysis and tensor algebra ~METHODS:Great a better kurnel that only works with tnsor algebra and creates the code to handle any zero entries in the data ~created a kurnal called taco that generates code to handle the zero entries in big-data sets ~
75 ~94 ~https://news.stanford.edu/2017/11/15/algorithm-outperforms-radiologists-diagnosing-pneumonia/ ~Stanford Researchers. ~Wanted to develop an algorithm that evaluates chest x-rays for signs of diseases. ~Researchers used a public dataset released by the National Institutes of Health Clinical Centre on 26th Sept. The dataset consisted of 112,120 frontal-view chest x-rays images labeled with 14 possible pathologies. ~None mentioned. ~The algorithm could diagnose all 14 of the pathologies labeled in the x-ray accurately. It also outperformed the four Stanford radiologists in diagnosing pneumonia accurately. ~
433 ~88 ~https://www.teslarati.com/openai-self-play-dota-2-musk/ https://blog.openai.com/dota-2/ ~The article is reported by Mike Tolzer. The project itself was undertaken by OpenAI, a non-profit artificial intelligence research company ~The goal of this project was to designed a bot for the popular video game DOTA 2. The aim was for this bot to learn how to play the game to such a degree that it could defeat the top-level players in the game, using self-play. ~OpenAI hypothesized that, for this project, supervised deep learning systems can only be as good as their training datasets. By using self-play as their data-set, the data would improved automatically as the agent gets better. ~OpenAI developed the bot in small, incremental steps. On March 1st, they began with classical reinforcement learning, where the bot 'kites' an enemy(Kiting is a strategy where a player with long-range attacks will fire at an enemy who has only melee-reach attacks, run away, fire, run away - repeating). Beyond simple strategies like this, OpenAI used only self-play with the bots. Agents undergoing self-play initially receive dense rewards for behaviours that aid exploration like standing and moving forward, which are eventually annealed to zero in favor of being rewarded for just winning and losing. The agents neural network is trained with Proximal Policy Optimization algorithms(PPO). This uses an adaptive Kullback-Leibler divergence penalty to control the change of the policy at each iteration. ~The results of the agent were tracked using TrueSkill. TrueSkill is a skill-based ranking system developed by Microsoft. The graph representing this is gradual increase from the bots inception in April, to its last record play in August. In the month of August, the bot had a string of victories against top players, often in 3-0 wins. On the tenth of August it was due to face the top player, Sumail. The agent defeated Sumail 6-0, and Sumail remarked that the agent was unbeatable. The following day the agent faced Dendi, an old crowd-favourite, and defeated him 2-0. ~While creating and having the agent learn was an impressive feat in itself, what is much more impressive here is the rate at which the agent learned, and went on to defeat the top players in the world. Large DOTA 2 tournaments are often played in huge football stadiums with sell-out crowds, and the agent was brought to tournaments like this. It often used strategies that impressed crowds to the extent it illicited cheers and loud reactions. What impresses me about this is that the agent gives the appearence that it is not just playing like an optimal-route machine, but an ingenious player that utilises unseen tactics and lightning reactions.   
48 ~67 ~https://www.eedesignit.com/star-the-smart-tissue-autonomous-robot/ ~STAR ~The Smart Tissue Autonomous Robot was created through the use of technologies designed to sense, process sensory information and perform actions for people with disabilities and seniors. ~Not Mentioned ~Not mentioned ~STAR is succesfully able to perform surgery but takes much longer than a doctor. ~
292 ~258 ~https://medium.com/@timanglade/how-hbos-silicon-valley-built-not-hotdog-with-mobile-tensorflow-keras-react-native-ef03260747f3#88fa https://www.forbes.com/sites/ianmorris/2017/05/16/this-hbo-silicon-valley-app-can-tell-hotdogs-from-not-hotdogs/#6e5888d37982 ~The source of this piece started as running gag on HBO's TV Show 'Silicon Valley'. The idea was an app called 'Not Hotdog'. The app would snap a picture of a piece food, and verify if it was a hotdog, or not a hotdog. The app was taken from fiction and made reality by a single developer involved in the show. While it may seem farcical at first, the overall process was a great example of how Machine Learning can be used. ~The goal here was to produce an app that could correctly identify a piece of food as either 'Hotdog' or 'Not Hotdog' ~The dataset used was quite varied. General images of hotdogs were used in varying situations - height/width, background, lighting conditions, cultural differences, perspective, composition, etc. Images that could attempt to 'fool' the neural network were also used - such as hamburgers, sandwiches etc. Aggressive data augmentation was used to counter the fact that not every image will be taken 'perfectly' eg Not every image will be snapped using a DSLR. The final dataset consisted of 150,000 images, of which 3000 were hotdogs.  ~The Keras implementation on TensorFlow was used for this project, along with Batch Normalization. This is is a technique to provide any layer in a Neural Network with inputs that are zero mean/unit variance. It activates of the previous layer at each batch, i.e. applies a transformation that maintains the mean activation close to 0 and the activation standard deviation close to 1. ~The agent is able to recognise the difference between hotdogs and non-hotdogs successfully. ~As previously mentioned, this exercise seems farcical and silly. However its implementation and extremely clear documentation provides an interesting exercise into the many applications of machine learning. 
73 ~99 ~http://www.digitaljournal.com/life/health/microsoft-enters-race-to-find-cancer-cure/article/506800 ~Microsoft ~To find a cure for cancers. ~"Biological data", such as cells, and pathways.   ~The first method is encapsulating the pathways of cells and the second is a tool called "Bio Model Analyzer". The tool allows ~The system has been used for pathways and "cellular stabilization". ~The attempts by Microsoft to tackle disease is interesting. However, I am sketpical that simply visualizing the data is the steps that are required.  
92 ~107 ~http://blog.kaggle.com/2017/04/26/dstl-satellite-imagery-competition-1st-place-winners-interview-kyle-lee/ ~Kyle Lee ~Detect and classify the types of objects found in satellite images. ~The images were 1km x 1km in dimensions and had 2 types, he 3-band images are the traditional RGB natural color images. The 16-band images contain spectral information by capturing wider wavelength channels. ~In summary his method was generate sliding windows, train a U-NET model and ensemble it with other models, and the applying special procedures to more detailed objects, such as waterways, cars and roads. ~Kyle was able to win the competition and was awarded $50000 ~
131 ~225 ~Article in Radiology Business titled "Deep learning proves effective in spotting liver masses in CT" http://www.radiologybusiness.com/topics/practice-management/quality/deep-learning-proves-effective-spotting-liver-masses-ct ~Koichiro Yasaka, MD, PhD Department of Radiology, University of Tokyo Hospital, Japan ~To enable the detection of liver masses or tumours on livers. ~100 liver mass image sets from CT scans in 2016, including 74 men and 26 women with the average age of 66 years old ~The article declares the use convolution neural networks were trained on CT scans to be able to detect tumours.  ~The article states that the model generated was successful at detecting liver masses without declaring success rate or accurate assessment. ~Despite not declaring the accuracy of the method in the article. A link to the study which states median accuracy for differential diagnosis of liver masses of 0.84. 
68 ~151 ~https://www.forbes.com/sites/tomdavenport/2017/11/05/revolutionizing-radiology-with-deep-learning-at-partners-healthcare-and-many-others/#6bb0d9665e13 ~Center for Clinical Data Science (CCDS) ~Improve computer-aided image based diagnoses using machine learning and other AI related techniques. ~Images from MRI scans. ~"Deep neural networks." Catastrophically nonspecific. Also, "a variety of other ML techniques." ~Assessment of deep-learning-based pulmonary nodule detectors showed that none of their outputs were the same: most focused on a particular aspect of the nodule that was not clear from the training. ~N/A. 
720 ~222 ~https://www.theverge.com/2017/8/9/16117850/deepmind-blizzard-starcraft-ai-toolset-api https://deepmind.com/documents/110/sc2le.pdf https://deepmind.com/blog/deepmind-and-blizzard-open-starcraft-ii-ai-research-environment/ ~The media article was covered by James Vincent of The Verge. The individuals involved in this project was a joined team assembled from employees at DeepMind and Blizzard Entertainment. DeepMind is a leader in artifical intelligence who look to push boundaries by developing systems to solve complex problems. This has often lead them to the area of gaming, both traditional and video gaming. The company made headlines worldwide with its program AlphaGo, which defeated a human Go player. AlphaGo then went on to defeat Lee Sedol, the current world champion. Their parters in this project, Blizzard Entertainment, is a video game company founded in 1991. They have several popular titles, 'Starcraft 2' being the one this project is focused on. Starcraft 2 is a 'Real-Time-Strategy' game, which focuses on strategy, unit control, and quick reactions. It is a title that is ideally suited for a project like this.  ~The Goal of this projet was to introduce the StarCraft 2 Learning Environment(SC2LE), a reinforcement learning environment based on Starcraft 2. This allows Starcraft 2 to function as research environment. DeepMind recgonised Starcraft 2 was a suitable environment to teach computers advanced skills like memory and planning, and as such created the environment for the benefit of AI researchers. The packaged includes an API, which is essentially the core of the project, that allows the AI to play the game feed back data to the researchers ~The data supplied by Blizzard Entertainment has been from 'replays' of previous games by human players. These fall into two tiers: Novice, and Grandmaster. The Novice matches are between two human players of average skill, while the Grandmaster matches are between players involved in e-sports in a profressional manner. Besides for replays of matches, Blizzard have also provided data in the form of mini-games that isolate certain gameplay elements like map exploration and resource collection, which can be used to develop a particular skill. ~The project makes use of a relative new and popular Reinforcement Learning Algorithm, the Asynchronous Advantage Actor-Critic (A3C) algorithm. A3C was created by DeepMind and is the faster, simpler, and more robust successor to the Deep-Q-Network - which was also created by DeepMind. As the name suggests, A3C is Asynchronous, and as such launchs several agents(workers) and lets them all interact with their own instance of the environment. The number of workers launched is relative to the host machines CPU, and the more that are launched, the more data that can be collected in a smaller window. ~In comparing agents trained on 'novice' data and 'grandmaster' data, it was found the agents on the lower spectrum were unable to develop a meaningful strategy for a full-game (Of length 30 minutes). The most succesful agent of this grouping was one that avoided constant losses by essentially running away for the length of the game (To expand on this point; One of the selectable teams in the game are called 'Terran'. Terran are humans from the distant future, and have the ability to let their buildings fly. To avoid defeat, the agent would uproot his buildings as the opponent approached, and fly them away to another location). The other group of agents - those who learned from the 'grandmaster' data converged to trivial strategies that only avoided distracting worker units from mining minerals, thus maintaining a stable improvement in the score(There are many metrics which dictate the overall score of a game on its complemention, and overall mineral collection is one of them). Thus, the score of most agents converged to simply preserving the initial mining process without building further units or structures. ~I think this a very interesting project and it does have a lot of potential to bringing interest in Machine Learning to a wider group of indivduals. Given that e-sports like StarCraft 2 are part of a multi-billion dollar industry, further projects in this field could aid the spread of Machine Learning. I had initially expected the 'Grandmaster' agents to form a near-optimal game, but I was incorrect. StarCraft 2, given its complexities, may not yet be suitable for this sort of learning but it is a step in the right direction certainly, and as previously mentioned the topic has sparked a lot of conversation on forums where the topic of Machine Learning is largely unheard of.  
234 ~88 ~https://www.theverge.com/2016/9/12/12886698/machine-learning-video-image-prediction-mit ~MIT Research dept ~To create a video from a still image, using machine learning. The idea being to emulate how humans can reasonably predict the next step a person will take or the movement of a bicycle, taking gravity and inertia etc into account without giving these predictions very much thought if at all. This project is considered to be a challenge in machine learning and machine vision.   ~2 Million videos from flickr were used as the training dataset in the themes of golf, beaches, train stations and hospitals (baby images). ~Flickr videos were stabilised to amend for camera shake. The footage was analysed to learn what the next behaviour might be in the image by guessing how the pixels might change.   ~Videos created are limited to a number of seconds, and the quality is poor, with blurred and fragmented looking moving that is generally far from realistic. Yet the article states that it's still an impressive feat of machine imagination, and another step toward computers that understand the world a little more like humans. The researchers acknowledge these limitations but are pleased that the motions themselves are plausible. ~It was also noted in the article that as humans we may predict the end of a movie half-way through, and that solving that problem with machine learning or any other set of tools will be more than a challenge. 
120 ~98 ~ Devex  https://www.devex.com/news/bringing-machine-learning-to-last-mile-health-challenges-91453 ~EasyScan  ~The goal of this algorithm is to detect malaria parasites in blood samples ~The machine learning algorithm was given healthy and unhealthy blood samples. ~ The machine learning algorithm analysed the blood smears and look for malaria parasite.   ~ The project was very successful and it is able to detect the malaria parasite in under 20 minutes. It also decreased the cost and removed a bottle neck in analysing a blood sample as a human doctor, who are in short supply, are no longer needed. ~ There is no description of the algorithm that they used but I believe that it was probably based off a difference from a "normal" blood sample 
178 ~80 ~https://www.techemergence.com/facebook-artificial-intelligence-hussein-mehanna/ ~Facebook ~To improve Facebook users' experience by detecting what posts they like to see in their newsfeed, and what posts they do not want to see. ~Assessed how often users interacted (positively or negatively) with different kinds of posts in their newsfeed. Also used Facebook's option to ban certain pages/posts from your feed to predict what other pages/posts that user would not like. ~Assessed how often users interacted (positively or negatively) with different kinds of posts in their newsfeed. Also used Facebook's option to ban certain pages/posts from your feed to predict what other pages/posts that user would not like. ~Facebook continues to be one of the largest and most succesful companies in the world, with over 2 billion active monthly users. This is in no small part due to the optimisation of the news feed, Facebook's main function (arguably aside from messenger). The team at Facebook seem to be very satisfied with what they have accomplished. ~As a long time Facebook user, I definitely have noticed the increase in quality of my news feed over time. 
51 ~436 ~Source:http://news.mit.edu/2017/artificial-intelligence-for-your-blind-spot-mit-csail-cornercameras-1009 AGENT:Adam Conner-Simons, self-driving cars GOAL:To reveal information about objects around corners DATA:No data  Methods: using cornercamers to get one dimension images of a hidden scene and taken multiple images and then stiching them together  Results: This results in finding the speend and trajectory of the object  Comments: 
210 ~96 ~http://www.earley.com/blog/lessons-alexa-artificial-intelligence-and-machine-learning-use-cases ~Seth Earley GOALS To use Amazon Echo, Amazon's cloud based intelligent agent, which interacts with voice recognition and using voice commands, performs tasks.  ~To set it up, Earley walked through some of its features such as asking about the weather, setting a timer, asking facts etc.  ~The usual functions worked well but it became quickly apparent that Alexa was not robust when questions were asked off script or there was even a tiny bit of variation in phrases said. When asked "How large are you?" or "What are your dimensions?", Alexa failed to understand. When asked "How tall is the Echo", Alexa was finally able to answer and other phrases like "How large is the Echo?" and "What are the dimensions of Echo?" were also successful. It turned out that adding the extra term "physical" to "dimensions" caused the algorithm to fail. ~To make Alexa work, Earley realised he had to use specific language to make the functions work. Even though he made a lot of mistakes, he eventually learned how to control the automation with his voice. ~I can see the benefits and excitement behind devices like Amazon Echo but I also understand that the importance of a good algorithm could make or break the product.  
92 ~239 ~https://thenextweb.com/artificial-intelligence/2017/11/13/alibabas-ai-is-the-blueprint-for-brick-and-mortar-stores-of-the-future/ https://www.technologyreview.com/s/609452/alibabas-ai-fashion-consultant-helps-achieve-record-setting-sales/ ~Alibaba. ~To provide fashion advice to in-store customers with a view to increasing offline sales. ~The inventory of the store the customer is in, and the items they have chosen to try on. Each item's tag is embedded with information, so there is no use of cameras. ~Deep learning. FashionAI is trained to recognise millions of items of clothing and give suggestions based on what other similar customers have bought, and according to the tastes of designers. ~Alibaba’s sales from Saturday’s Singles Day event exceeded 25 billion dollars. ~No comment. 
118 ~143 ~https://www.forbes.com/sites/forbesagencycouncil/2017/11/15/how-machine-learning-can-maximize-the-success-of-marketing-campaigns/#6127bdcf65c7 ~Marketing departments ~To revolutionise marketing by introducing machine learning to analyse massive data on customer behaviour and identify trends and create advertisements which emotionally connect with users.  ~Customer data from surveys, social media, reviews, public opinion etc ~Marketing departments want to use lexical analysis algorithms which learn from the ways humans react to different language and various ways of expressing the same idea. ~Marketing campaigns which maximise emotional connection with the consumer in order to increase the chances of them purchasing the product.  ~This idea has been around for a long time, however, it is only starting to be employed on a wide basis as machine learning becomes accessible to even medium and small businesses.  
543 ~52 ~https://www.wired.com/2014/01/how-to-hack-okcupid/  ~The article was reported by Kevin Poulsen, while the learning in question was done by Chris McKinlay. McKinlay, at the time, was a mathematician undertaking his PhD. The title of his thesis was 'Large-scale data processing and parallel numerical methods'.  ~McKinlay was an avid user of dating websites, with limited success. In 2012, McKinlay was working on machine code, when he had the revelation that he could he was approaching one particular dating site - OkCupid - in the wrong manner. McKinlay decided to tackle the website in a mathematical manner, with the goal of boosting his successes on the website.  ~OkCupid at its heart is mathematical, having been found by Math majors from Harvard, and uses a computational approach to matchmaking. Upon signing up, and in regular intervals, users are asked multiple-choice questions on a range of data - politics, religion, family, general outlooks. On average a user will answer 350 questions, and OkCupids matching algorithms uses this data to calculate two peoples compatibility. The closer to 100% the match is, the more likely the two parties will interact well. In order to gather suitable data, McKinlay wrote a Python script. The script would use 12 fake OkCupid accounts to scrape all available information off their profiles, including ethnicity, height, smoker or nonsmoker, and astrological sign. Furthermore, McKinlay gathered information about the womens responses to their own question. After overcoming some data-mining hurdles, McKinlay had gathered 6 million questions and answers from 20,000 women across the USA - after 3 weeks.  ~McKinlay used the K-Modes algorithm to accurately survey the data. While K-Means is used for numerical data, k-modes came out of a paper in 1998 by Zhexue Huang and was designed with categorical data in mind. Instead of distances seen in K-Means, it uses dissimilarities(the quantification of the total mismatches between two objects - the smaller this number, the more similar the two objects). Furthermore, instead of using means, it uses modes. As with k-means, K-Modes will converge into suitable clusters. After McKinlay had crunched the data, he selected the cluster that suited him most - in this case it was women in their twenties, who looked like musicians and artistsm, and another cluster for slightly older women who had professional creative jobs. Following this, McKinlay created a profile and filled out 500 questions that were popular with both clusters. While he answered the questions honestly, he left the importance(users not only answer a question, but rate its importance to them personally) up to another algorithm called adaptive boosting. ~The experiment was without a doubt a mathematical success, with McKinlay getting 400 visits to his profile daily, and recieved a constant stream of messages. McKinlay did succeed in getting many dates, and as the summer came to closed he had been on 55, thanks to the his use of algorithms. Eventually McKinlays uses of machine learning algorithms found him a match, Tien Wang, to whom he is now engaged.  ~Though the article may seem somewhat trivial or maybe a little silly, I think it is a very practical and approachable piece on a practical use of machine learning. McKinlay used well known algorithms to great effect and ultimately 'cheated' the system and completed his goal. 
365 ~261 ~ http://im2recipe.csail.mit.edu/ </br> http://www.wired.co.uk/article/ai-food-scan-images-pic2recipe https://www.digitaltrends.com/photography/pics2recipe-mit-research/ http://news.mit.edu/2017/artificial-intelligence-suggests-recipes-based-on-food-photos-0720 ~ Pic2Recipe is a website created by Nick Hynes (electrical engineering and computer science graduate from MIT) and researchers from CSAIL at MIT.  ~ It consists of a neural network that has been trained to predict the ingredients in a dish based on a picture of that dish. "Its training allows it to find patterns and make connections between the food images and the corresponding ingredients and recipes". ~ The work is built off datasets including "The Food-101 Data Set" and a database from the City University in Hong Kong. Pic2Recipe uses training data from websites like All Recipes and Food.com, entitling the set "Recipe1M". This training set contains over 1 million recipes that are annotated with information about recipes and ingredients of a wide range of dishes.  ~ The neural network was then trained to find patterns and connections between images in Recipe1M. ~ Although Pic2Recipe proved to be quite good at performing its duties, it sometimes failed in identifying differences in obscure food items such as soups and smoothies. It also struggled with food items that had more than one recipe possibility.  Minor  adjustments were made to accommodate for a food item with multiple recipe options. This was done by cross-referencing ingredients, before continuing to check each possibly recipe. Future plans for Pic2Recipe include functionality that can determine the method of cooking from an image, and functionality that can specify an exact ingredient as opposed to a general class e.g. distinguish between a cherry tomato and a grape tomato. ~ The idea of computer vision and neural networks combining to create something that can analyse an image to such an extent, is quite an exciting one. The applications of this work could translate into a multiple of possibilities. The refining of Pic2Recipe and expansion on Recipe1M could make for big advances in the analysis of nutrition. From this we could learn about peoples' cooking and eating habits, standards in cooking and possibly learn from food in a cultural context. Tracking personal nutrition could also be a possible outcome; developing a mean of determining the nutrients and calories in a meal from an image. 
97 ~115 ~Unsupervised Learning of Disease Progression Models http://people.csail.mit.edu/dsontag/papers/WanSonWan_kdd14.pdf ~Xiang Wang, David Sontag and Fei Wang ~Modeling disease progression based on real-world evidence is a very challenging task due to the incompleteness and irregularity of the observations, as well as the heterogeneity of the patient conditions. ~Patient records (that were diagnosed with COPD) ~Expectation Maximization (EM) based algorithm. ~They trained the model using the exact same settings but without the anchors. As a result, conditions from different comorbidities were mixed into one. ~The techniques presented in this paper open the door to many new applications of disease progression modeling. 
248 ~78 ~ https://link-springer-com.jproxy.nuim.ie/article/10.1007%2Fs10705-017-9870-x ~ Sub Sahara African agencies ~ To produce a map of detailed spatial prediction methods on 15 soil nutrient content in Sub Sahara Africa. (SSA) Such information could be used to support agricultural development in an area where over 50% of the world's potential land is available for cultivation yet only 9% is arable land and 1% permanently cultivated. ~ Input data of soil samples from 59,000 unique locations taken over a measured period of time, and measured depths from SSA countries including but not exclusively, Uganda, Rwanda, Burundi and Ethiopia. These samples were provided by seven different agencies. Nutrients observed included for example organic carbon (C) and total nitrogen (N), total phosphorus (P), and extractable: phosphorus (P), potassium (K), calcium (Ca) etc. Macro nutrients and micro nutrients were Observed. ~ For model fitting and prediction the researchers used Random Forest and Gradient Bosting Tree Machine Learning Methods. ~It was found that nutrients S, P, and B tended to be more difficult to model spatially using their framework. But significant models CAN be produced for most targeted nutrients. And maps of micro/macro nutrients could be used to mark areas of nutrient deficiency and sufficiency with a view to cropping and nutrient adjustments. Cluster analysis showed that there was potential for creating 20 zones in SSA requiring similar nutrient management.  ~Some areas in the SSA area were under-represented in terms of soil samples. Collection of more samples from these areas would go towards improving prediction accuracy. 
167 ~164 ~Article by UCL news titled "Improving clinical trials with machine learning" https://www.ucl.ac.uk/news/news-articles/1117/151117-machine-learning-clinical-trials/ ~Dr Parashkev Nachev, UCL Tianbo Xu, UCL Professor Geraint Rees, UCL ~The aim of the study was to improve the ability to determine whether a new drug works in the brain. ~Gaze direction, objectively measured from the eyes as seen on head CT scans upon hospital admission and from MRI scans typically done 1-3 days later. ~The specific machine learning technique was not described simply referring it to it as  "high-dimensional machine learning methods" ~Their technique is reported to be able to determine if drug impact is significant when the lesion was shrunk by only 55% as opposed to 78.4% by conventional methods. ~This a low detail article which appears to more describe the hopefull ambitions of the use of machine learning as opposed to the current application. Although no specific machine learning method was mentioned it is likely that convolution neural networks were used as they were working with CT scan image data.  
352 ~147 ~This article was written by Arthur Juliani and published by Unity. https://blogs.unity3d.com/2017/09/19/introducing-unity-machine-learning-agents/ ~This project was undertaken by Arthur Juliani and the machine learning team at Unity. ~To design a system that provides greater flexibility and ease-of-use to the growing groups interested in applying machine learning to developing intelligent agents, specifically for academic research in complex multi-agent behaviour, industry research in large-scale paralle training regimes for robotics and game developers interested in filling virtual worlds. ~Unity uses three main kinds of objects within learning environments. These are Agent, Brain and Academy. These three objects work together to improve the intelligence of the agents. ~Unity agents have a variety of possible scenarios which depend on how agents, brains and rewards are connected. A single agent is linked to a single brain and is the traditional way of training an agent. An example is any single-player game, such as Chicken. A Simultaneous Single-Agent consists of multiple independent agents with independent reward functions linked to a single brain. An example would be training a dozen robot-arms to each open a door simultaneously. Adversarial Self-Play contains two interacting agents with inverse reward functions linked to a single brain. A Cooperative Multi-Agent has multiple agents with a shared reward function, linked to either a single or multiple different brains. A Competitive Multi-Agent has multiple interacting agents with inverse reward funcion linked to either a single or multiple different brains. And finally, an Ecosystem which has multiple interacting agents with independent reward functions linked to either a single brain or multiple different brains. ~With the release of this open beta SDK, Unity has provided developers and researchers with the abilty to transform games and simulations created using the Unity Editor into environments where intelligent agents can be trained using Deep Reinforcement Learning, Evolutionary Strategies, or other machine learning methods through a simple to use Python API. ~I feel this beta is a fantastic release from Unity as it provides researchers and developers with a very powerful tool in the form of open-source software, as well as examples to get them started with machine learning in research and game development. 
54 ~134 ~Huffington Post https://www.huffingtonpost.com/entry/three-real-use-cases-of-machine-learning-in-business_us_593a0e91e4b014ae8c69df37 ~ZenDesk ~They wanted to look for a solution to better target people who want to buy their products. ~Users online ~They identified patterns in their contact data and used it to create patterns of persona. ~They succesfully achieved their goal. ~This has many uses and could potentially lead to increased users. 
187 ~32 ~https://blog.openai.com/dota-2/ ~Open AI ~Wanted to create a machine that could beat professional DOTA players at their own game. ~Had the machine play against numerous pro players and recorded the results. ~Had the machine constantly play 1v1 against itself; learning what strategies led to wins more often that not. Once ready, moved on to playing against real people, including many professional players. ~The machine absolutely destroyed almost all pro players that it was put against in a 1v1 match. The next goal of the Open AI team is to build this machine into one that can control 5 characters, and beat a full professional team in a standard 5v5 match. ~I was legitimately amazed at how well the bot performed. Games like DOTA are extremely skill intensive. Both in a mechanical sense (accuaracy, dodging etc.) and a knowledge sense (decision making, predictions etc.). The fact that the bot managed to teach itself these skills so well that the best DOTA players in the world couldn't beat it is genuinely insane. I look forward to seeing if they can manage to win a 5v5 match; a significantly harder challenge. 
83 ~63 ~ Standford https://stanfordmlgroup.github.io/projects/chexnet/ ~ Stanford  ~ To use Xray images to automatically detect if a patient has pneumonia ~ ChestX-ray14(the largest public database of X-rays): they used a series of chest X-rays some with and some without pneumonia ~ CheXNet, is a 121-layer convolutional neural network that inputs a chest X-ray image and outputs the probability of pneumonia along with a heatmap localizing the areas of the image most indicative of pneumonia ~ can detect pneumonia better than most radiologists  ~
72 ~167 ~Facebook Research: Cultural Diffusion and Trends in Facebook Photographs (https://research.fb.com/publications/cultural-diffusion-and-trends-in-facebook-photographs/) ~Quenzeng You, Dario Garcia, Manohar Paluri, Jiebo Luo, Jungseock Joo ~Identifying location and time related trends for pictures post on Facebook. ~250 millions de-identified photographs with upload time and location. ~ResNet-50 (deep residual network) ~Automated classification of visual content of photographs in social media is an effective means to assess the local and global trends of various cultural activities and lifestyles. ~
69 ~127 ~HUFFPOST https://www.huffingtonpost.com/entry/three-real-use-cases-of-machine-learning-in-business_us_593a0e91e4b014ae8c69df37 ~ZenDesk ~Lokking for a solution to better target audiences ready to purchase their products. They felt their audience was too broad and led to excess costs for pay-per-click and search angine marketing leads. ~No data mentioned ~Used MarianaIQ's social media engagement platform to help create categories of personas. ~ZenDesk claims that its lead volume increased by a multiple of four and effectively drove down cost-per-lead. ~No comments 
320 ~174 ~This article was written by Sophia Ciocca and published by hackernoon. https://hackernoon.com/spotifys-discover-weekly-how-machine-learning-finds-your-new-music-19a41ab76efe ~The models mentioned in this article are used by Spotify. ~To create a new personalised playlist of 30 songs every week, which have not yet been listened to by the user, which they will probably be interested in hearing. These playlists are personal to each of Spotify's over 100 million users. ~Spotify uses implicit feedback (specifically stream counts of tracks listened to) as well as additional streaming data such as saved tracks or if the artist page was visited after listening. They also use track metadata, news articles, blogs, and other text around the internet as well as raw audio data. ~Ciocca researched Spotify's three types of recommendation models which are employed by the company: Collaborative Filtering, Natural Language processing and Raw Audio Models. With Collaborative Filtering, Spotify uses a matrix to represent users as well as songs. Using a matrix factorization formula, the matrix can split users and songs into seperate vectors, which can then be compared to with each other to find similarities. With Natural language processing, Spotify crawls the web for written texts about music and figures out what is being said about particular artistsand songs. Again this data is stored in vectors for comparison. With Raw Audio Models, spotify uses convolutional neural networks to compute statistics of the learned features across the time of the song. This data allows for even more comparison for similarities in songs. ~By combining these three models, Spotify is able to create a uniquely powerful discovery engine, which can ultimately recommend new music to their users which has a higher accuracy of being enjoyed by that user on a personal level. ~I think this is a great approach by Spotify, as it creates the feeling of a personal touch, with care provided to each of its users, which in turn will keep them loyal and subscribed to the service. 
55 ~109 ~https://www.newscientist.com/article/2147472-ai-spots-alzheimers-brain-changes-years-before-symptoms-emerge/ ~Nicola Amoroso and Marianna La Rocca ~using a machine learning algorithm to discern structural changes in the brain caused by Alzheimer’s disease so that it can identify changes in the brains of people likely to get Alzheimer’s disease almost a decade before doctors can diagnose the disease from symptoms alone. ~none ~none ~none ~none 
204 ~57 ~IEEEXplore. http://ieeexplore.ieee.org/document/7870220/ ~Pace University. Researchers: Javid Maghsoudi, Charles C. Tappert, Seidenberg Schcool of CSIS, Pace University, Pleasantville ~To identify and authenticate the user of a smartphone using behavioral biometrics such as gait, grip, touch gestures ~60 people, 10 using each of 6 different phones, were used to collect the data.  Five of the phones had both accelerometer and gyroscope sensors and one had only an accelerometer.  1200 distinct trials were carried out resulting in 2200 different datasets.   ~A simple division of the trials was carried out and the motions and pauses were identified, using advanced feature extraction, and put in separate segments.  Several algorithms were used to identify the user of a phone.  These included SVM, Naive-Bayes, K-Nearest Neighbours and a Multilayer Perceptron.  10-fold cross validation was used to evaluate the algorithms with one tenth being held for testing. ~The best results were achieved from using 2 sensors together with complex extraction and either the SVM or multilayer perceptron algorithms ~Each of trials was carried out in a single period of time.  The researchers suggest that it would be interesting to see if the results would be affected by running the trials over multiple time periods. 
138 ~217 ~The Verge - Salesforce created an algorithm that automatically summarizes text using machine learning https://www.theverge.com/2017/5/14/15637588/salesforce-algorithm-automatically-summarizes-text-machine-learning-ai ~Andrew Liptak & Salesforce ~Saleforce is a cloud service provider which developed a Machine Learning algorithm which will summarize text, making it more concise for the reader. ~Large data set of sentences along with their more concise versions. METHOD Salesforces algorithm scans user written text and using the training data it was provided with aims to provide a shorter, but still readable, version of the text. ~Salesforce has a much success with this algorithm which they trained to not only understand words in the sentence but also allowed it to introduce new words. They are now offering this service to their customers. ~This article praises Salesforces algorithm as a major timesaver in the lives of many customers as way of digesting information quickly. 
111 ~112 ~https://www.technologynetworks.com/tn/news/scientists-use-machine-learning-to-analyze-language-in-movies-294179 ~A team at the University of Washington ~To create a system that could quantify the power and agency of characters in movies based on their dialogue and stage  directions. ~The scripts of nearly 800 movies. ~Built on a previous project involving "connotation frames" but doesn't mention algorithms employed. Algorithms cross  referenced words with a bank of 2000 commonly used words whos connotations were scored on both their power and agency. ~Found gender bias in the writing of many characters. For example in "Black Swan" the female leads dialogue gives her more  agency but in the stage directions more power and agency was given to male characters. ~
393 ~192 ~Louis Colombus, Machine Learning Is Redefining the Enterprise In 2016 https://www.forbes.com/sites/louiscolumbus/2016/06/04/machine-learning-is-redefining-the-enterprise-in-2016/#3f93c43f1871 ~Forbes, june,2016 ~Using historical data, ML will help businesses grow by basing their decisions on business models with the best outcome.  ML will lead businesses into environs with minimum human error and stronger cyber security.ML will completely enable new interactions between customers and companies and eventually allow true intelligent enterprise. Most of the global enterprises have cantered their corporate strategy on digital transformation.  Machine learning is a major part of this transformation. Applications, Algorithms and frameworks learn from a huge amount of structured and unstructured data such as Text, image, video, voice, body language and facial expression.  This creates room for a wider dimension of applications and algorithms ranging from healthcare to video games and self driving cars.  ~Enterprises use historical data within their coppany to train and algorithms and predict future outcomes. ~While machines can automate lengthy and repetitive tasks, they can also be used to predict the outcome of new data.  This predicted data is used to make decisions. Companies rely on these predictions to advance their businesses and reduce the risk taken in decision making as it minimises human error.  Prediction algorithms rely heavily on behaviour signals.  The more the user shares with the algorithm the better the prediction becomes, one example is streaming platforms such as Netflix and Amazon prime where, depending on the users behaviour, the platform can recommend and advertise content.  These algorithms define a strategy to drive desired sales and market outcomes. Business processes will also become automated and as the business changes these algorithms will constantly update and adapt themselves. These business processes range across a wide spectrum of contract management, customer service, finance, legal, quality assurance, pricing and production.  ~ML in is changing enterprises for the better according to most companies who are applying machine learning to sales and marketing.  With machine learning, companies are gaining a greater predictive accuracy.  According to the Accenture institute of higher performance, 76% of the companies that use machine learning claimed to be targeting a higher sales growth. ~with the increase in the use of machine learning within enterprises, a rise in competition between the companies is also present. This forces companies to consider advancing their technologies to increase sales as well as predict future outcomes. 
264 ~135 ~By Hyeokjun Choe, Seil Lee, Hyunha Nam, Seongsik Park, Seijoon Kim, Eui-Young Chung, Sungroh Yoon url:https://arxiv.org/abs/1610.02273 ~Cornell University Library. The library stands  at the center of intellectual life  for the university. Expert librarians help navigate world-class collections to help assist with papers, exam prepations and long term projects. This library is a partner in study, research and teaching for the university. ~To augment memory or storage with processing power. This would increase the potential for accelerating computing and would greatly help in reducing the power requirements for computation in computers. To have more advancement in the field of progressing in accelerating processing power by augmenting memory or storage in computers.  ~The possible improvements of making further advancements in accelerating computing in this paper was evaluated for machine learning using a new platform they have developed, based on machine learning workloads. ~NDP-Near Data Processing ISP-Instorage Processing SSD- solid-state drives Instorage processing (ISP) of machine learning workloads were simulated to evaluate the potentail for their advancements in accelerating computation using NDP. ~By using near-data processing (NDP), the authors discovered that they could augment memory or storage space to accelerate computation. This would also result in reducing the amount of power computer would require in order to perform certain task as NDP would make the computer more efficient. ~I found it quite interesting as to how the methods the authors had developed and discovered to have the potential to significantly improve computation among CPU's globally in making them more memory efficient, ECO-friendly, faster and more powerful by using the method of near-data processing (NDP). 
78 ~59 ~https://which-50.com/using-machine-learning-detect-autism/ ~Cognoa ~using machine learning to detect cognitive disorders in children up to 13 months earlier than traditional diagnosis methods ~very large numbers of historical patient records ~use data to capture incredibly subtle patterns that might indicate the presence of cognitive disorders ~Earlier this year Cognoa secured $US 11 million in funding, taking their total funding to $US20.4 million. According to Abbas, Cognoa has already helped 300,000 parents. ~Another life saving and incredibly useful application of machine learning 
288 ~316 ~VisualDx Website: https://www.visualdx.com Web article: http://uk.businessinsider.com/visualdx-machine-learning-app-for-skin-diagnosis-ceo-interview-2017-11?r=US&IR=T/#visualdx-is-intended-for-use-by-doctors-to-confirm-and-validate-diagnoses-it-allows-doctors-to-search-by-symptoms-signs-and-other-patient-factors-1 ~VisualDx, a visual clinical decision support system. ~The image that the person takes of their skin and compare it against their Medical Image library, which is in fact the world's largest medical image library. The database has about 32,000 images to train its models, scanned from old collections of slides and films. METHOD The method was not explicitly mentioned but they use a smartphone camera along with a ML model called CoreML used to automate skin image analysis to assist in dermatologists with diagnoses. Instead of uploading the photos the person takes online to a server for processing, with the use of CoreML, runs the algorithms on your phone. CoreML was announced by Apple. The image is analysed on the phone and never has to go to the cloud. The model is an identification neural network and with it, it is trained by researchers at VisualDx, using their own library of medical images.   ~The result of this is that doctors can use this to confirm and validate diagnoses, as well as allow doctors to search symptoms, signs, and other patients' factors.  ~An interesting comment from the CEO of VisualDx, Art Papier, was how he said that  automatic diagnosis wasn't going to happen soon, even considering the achievements they made with these models and applications of Core ML and their app. He said that the "hype circle" of machine learning is "off the charts" and how it'll take a lot more to get there. Papier said "Machine learning will get you into a category on this, to get to the final mile, you have to ask the patient did you take a drug a week ago. Did you travel".  
196 ~142 ~ https://www.forbes.com/sites/bernardmarr/2016/12/29/4-amazing-ways-facebook-uses-deep-learning-to-learn-everything-about-you/2/#158f79d63090 ~ FACEBOOK ~ To recognise people in photos and use this to suggest friends or suggest who to tag in the photo ~ Pictures uploaded to facebook of themselves and pictures people upload containing other people. ~ Facebook employs a deep learning app they have developed called DeepFace which learns to recognise people in photos. This neural network is fed all the available pictures of the user to generate a set of values corresponding to that person, these are then used to identify these people in other pictures that their friends perhaps may have shared. ~ The neural network was remakably successful being regarded as one of the most advanced image recognition tools beating humans in extensive testing where the goal was to discern if 2 different images are of the same person. The success rate for DeepFace was 97% and the success rate for humans was 96% ~ This was a very successful bit of software but it got removed in 2013 from European accounts due to issues with infringing on privacy where people argued that with high level images of a crowd DeepFace could start putting names to faces it can see. 
76 ~99 ~http://in.pcmag.com/google-maps-for-mobile/114355/news/google-maps-gets-a-dose-of-machine-learning ~Google ~Wanted to to use photographic data to extract information from street signs to map new addresses. ~Uses photographic data that Street View cars gather.  So far they have collected more than 80 billion high resolution images  ~None mentioned. ~Google have automated the process of reading content, such as street names and numbers, from the images collected by Street View. The latest algorithm achieved an 84.2 accuracy on French street name signs. ~
205 ~80 ~https://9to5mac.com/2017/11/16/apple-machine-learning-journal-facial-detection/ ~Apple ~To implement facial recognition for the purposes of phone unlocking, photo categorising etc while maintaining a high degree of user privacy. ~Captured data from the user during the setup process, data from the user's photo library, new captured data every time the user attempts to unlock the phone using facial recognition. ~Deep learning framework which protects user privacy and has been written to run efficiently on a handheld device. The algorithm has been written in such a way that it can analyse an entire photo library in a reasonable amount of time while sharing resources with other applications and the OS on phone. The algorithm runs entirely on device as photos which are backed up to the cloud are encrypted first for user privacy.  ~The framework is able to recognise faces in photos and group photos on the device by the people in them and is also available to third party applications while maintaining user privacy. The framework is also being used for facial unlocking. ~Apple has created a sophisticated algorithm for facial recognition which maintains their company policy of protecting user data while also managing to be powerful, available in a limited fashion to third parties and execute with incredible efficiency.  
138 ~27 ~http://www.robots.ox.ac.uk/~parg/pubs/mann_maxent_final.pdf http://www.gaussianprocess.org/gpml/chapters/RW.pdf ~A cross-university multi-discipline team. Members came from Dept. Engineering Science & Dept. of Zoology, Oxford, Microsoft Research, Cambridge and Dept. of Zoology University of Sheffield. ~Prediction of Homing Pigeon flight trajectories. ~Recorded flight GPS data was used train the model. 31 birds, each with over 20 test flights from different test sites. ~Used the Gaussian procces Machine learning Model. Took each flight and using equations outlined in the paper, predicted its immediate successor. This was then compared against a prior computed flight path probabilty which gave it a score (Marginal Information Gain). Model was then retrained. ~As number of flights increased, M.I.G increased drastically. After 20 flights, future flight paths were extremely predictable. Further application for this would be to apply to other animal movement paths. Eg. locating food sources in the ocean by tracking seagulls. 
143 ~102 ~Online article on ZDNet http://www.zdnet.com/article/zillow-machine-learning-and-data-in-real-estate/ ~Zillow, a real estate company ~To estimate or predict property values using machine learning. ~Various, enormous data with regards to homeowners who buy and sell houses or properties in different locations (cities, regions, etc.). ~Linear regression, decision tree, classification techniques. ~The results have been relatively successful. Zillow's Zestimate has grown to something that affects the real estate market since the output would affect many buyers and sellers, who plan to buy or sell properties.  ~I read about linear regression techniques when it came to real estate prices. The simplest example was a table with two features. I can only imagine the implementation of Zestimate: big data manipulation, advanced machine learning concepts. I have no doubt that Zestimate has the push or the pull to affect customers' decisions with regards to buying or selling properties in the market. 
150 ~41 ~Story 6: https://blog.openai.com/dota-2/ ~Developed by OpenAI in conjunction with Tesla. ~To create a bot capable of beating top level players in the game Dota 2. ~Basic ruleset from the source code of Dota 2. ~A basic bot was developed initially, and was taught the basic rules. Then, this bot was made to play against itself thousands of times, while making small, random adjustments to play style each game. The method used was very similar to the genetic algorithm, in that these random adjustments were kept if they were shown to improve results. ~The bot eventually trained itself to be able to beat any player. ~It is worth noting that the bot was only developed for a sub-type game, as opposed to the full game. A normal game of Dota 2 consists of a 5 v 5 match up, where this bot was trained only for 1 v 1 match ups. 
216 ~168 ~PsyPost, NCBI. http://www.psypost.org/2016/07/machine-learning-puts-new-lens-autism-screening-diagnostics-43778,  https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4958551/ ~Researchers:  Daniel Bone, Matthey Black, Shirkanth Naravanan; University of Souther California, Somer Bishop; University of California, Matthew Goodwin; Northeastern University, Catherine Lord; Center for Autism and the Developing Brain, Weill Cornell Medical College ~To improve on existing autism spectrum disorder (ASD) diagnostic and screening tools.  ~The data used consisted of the test scores of 1264 people with ASD and 462 with non-ADS disorders.  The test results were obtained using the Autusm Diagnostic Interview-Revised (ADI-R) and Social Reponsiveness Scale (SRS) tests.  ~A support vector machine (SVM) was used to create the algorithms.  A multi-level cross-validation was used to test the algorithms performance.  A second layer of cross-validation of each training set was used to tune parameters and select features.  The first layer used 5-fold cross-validation and the second layer used 3-fold cross-validation on the training data for the first-layer.  ~Algorithms were able to reach 89% specificity for individuals under 10 years old and 87% for those above 10.  They also enabled identification of redundancies in the questions asked during the tests and identify five ADI-R questions that were able to keep 95% of instrument performance. ~Data was only included for verbal individuals due to the limited amount of data available for nonverbal individuals 
216 ~131 ~https://thenextweb.com/artificial-intelligence/2017/11/15/google-brings-on-device-machine-learning-to-mobile-with-tensorflow-lite/ ~Google ~To bring a machine learning framework to mobile devices for developers which can run efficiently and without cloud connectivity. ~Developer-provided data allows the machine learning algorithms and models to run locally on the device and result in outcomes. ~Google's new TensorFlow Lite framework is available for preview on iOS and Android and allows developers to provide a machine learning model and data and result in outcomes on the devices. The framework does not access the cloud, but runs offline.  ~The framework is applicable for a wide number of use cases, such as interacting with smart appliances in the user's home and housing intelligence for these appliances on the device itself, rather than cloud computing providing the intelligence. It also allows offline uses for apps, appliances etc, so intelligent machine learning algorithms can run on apps and in the user's home etc even if they have no internet connectivity. It also strengthens user privacy as all data is held locally. ~This is an excellent innovation by Google which will allow smart devices and the like to continue to work intelligently even when the user has no internet, creating a world in which smart appliances are available 24/7, even in situations where the internet is down, severing a reliance on external technologies at least temporarily.  
77 ~71 ~https://www.siliconrepublic.com/machines/machine-translation-adapt-dcu ~Adapt Centre’s MT team at Dublin City University. ~Language Translation. ~Source Sentence (English language sentence), human provided translation ~Neural Machine Translation ~Using one language as a pivot language between the translation of 2 other languages. Example; lots of Engish-Arabic/English-Greek translation data, not much Arabic-Greek translation data. Use English as a pivot language for translation from Arabic to Greek and vice versa. ~Interesting use of MT in disaster response situations where translation may not be readily available. 
291 ~408 ~ Washington Post website: https://www.washingtonpost.com/news/innovations/wp/2016/03/10/googles-psychedelic-paint-brush-raises-the-oldest-question-in-art/?utm_term=.c0fc2c081a5b Fortune.com web article: http://fortune.com/2016/03/01/google-deepdream-art/ DeepDream website: https://deepdreamgenerator.com Wikipedia page that explains process of the DeepDream program: https://en.wikipedia.org/wiki/DeepDream ~ DeepDream, a company that made an app that generates artistic interpretations of everyday life, real images. ~Create a machine that takes images, interprets them and makes interprets them artistically with different styles, etc. ~They used images, one that had artistic stylings and one real-life image to morph together. METHOD The method wasn't explicitly explained, only mentioning on their website that they used different AI algorithms to generate visual content merging image styles with real-life images and content.  The software detects faces and other patterns in images and classifying them. The method of running it in reverse after being trained resembles back propagation optimisation but instead of adjusting the weight heights, the values don't change but the input is adjusted. Also with the use of gradient descent, may result in the input producing "images in which adjacent images have little relation [...] thus the image has too much high frequency information". This is solved with the use of a "regulariser" which chooses inputs that are "simply smooth". All this occurs by adding "dreamed" inputs to the training set and can improve the training times. ~The results are abstract images created by the machine by fusing the images given as input and creating abstract images that were used in apps to allow people to transform their very own images and generate merging content and image styles. ~It is interesting the way machines and software that have been created are challenging us in the very things we see like art and bring a new outlook to the way the mid works and another demonstration of the power of neural networks. 
168 ~126 ~National Institutes of Health https://www.nih.gov/news-events/nih-research-matters/machine-learning-identifies-suicidal-youth ~A research team led by Dr. Marcel Just at Carnegie Mellon University and Dr. David Brent at the University of Pittsburgh. ~Using machine learning to analyse brain images and identify people with suicidal thoughts. ~fMri data from 33 participants to train a machine-learning system to spot differences in networks of brain activity. They then tested it on brain images from 34 participants. ~Supervised machine learning. ~The scientists used data from 33 participants to train a machine-learning system to spot differences in networks of brain activity. They then tested it on brain images from 34 participants. The system correctly identified 15 of 17 suicidal people and 16 of 17 controls with an accuracy of 91%. The system was also able to distinguish the 9 who had previously made a suicide attempt from the 8 who hadn't with an accuracy of 94% (16 out of 17). ~Being able to detect suicidal tendencies by showing words to a person shows promise for better mental health treatments. 
87 ~89 ~https://futurism.com/a-new-alexa-like-system-is-helping-robots-understand-context-clues/ ~Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL). ~To help robots understand context with a view to processing complex user requests. ~The robot's environment ~Computer vision methods to help the robot to visualise its environment, followed by periodically storing their observations for future reference. This "episodic memory" can then be reviewed when a request is being processed. ~When ComText was testing using a two armed robot (Baxter), the robot was able to perform 90% of the complex commands asked of it correctly. ~No comment. 
56 ~104 ~Source:https://www.nature.com/news/how-machine-learning-could-help-to-improve-climate-forecasts-1.22503 ~GOAL:use machine leaning to help improve climate forecasts ~Methods:Use machine leaning to first go through the data available to create models of climate change and secondly use macine learning to create algoritms to predict forest fires	 Results:creates models of the climate and continue to do so better models are formed and make improved forecasts Comments: 
56 ~174 ~Spotify's Discover Weekly: How machine learning finds your new music (https://hackernoon.com/spotifys-discover-weekly-how-machine-learning-finds-your-new-music-19a41ab76efe) ~Spotify ~Present to the user a generated playlist corresponding to there music taste. ~Raw audio model, metadata and search terms. ~CNNs ~Adding a CNNs layer to the recommendation system reduce the discrimination of less known or popular tracks and allow discovering of more tracks. ~
147 ~108 ~http://blog.kaggle.com/2017/04/20/dogs-vs-cats-redux-playground-competition-3rd-place-interview-marco-lugo/ ~Marco Lugo ~Predict wheter the a given image in the dataset is a cat or a dog ~25000 images of cats and dogs were provided for users to train their model and another 12500 for testing. ~Marco used an ensemble of models, taking a weighted average of the results of VGG16, ResNet50s, Xception among others, to produce a final result. ~While he finished 3rd in the competition, he highlights a couple of interesting facts. One was that even adding poor performing models into the mix, yielded positive results for the final ensemble. The second remark is the difference in results he had when changing from standard ReLU to Leaky ReLU and Randomised Leaky ReLU. ~This seems like a good challenge to demonstrate the capabilities of machine learning, but also its limitations. I wonder what would the model output if we try to classify horses or zebras. 
255 ~82 ~ https://www.ucl.ac.uk/news/news-articles/0816/110816-my-text-in-your-handwriting ~UCL Computer Scientists UK ~ To take a paragraph of handwriting, and produce text in that handwriting. It is referred to as 'My Text in Your Handwriting'. Proposed applications include for example, stroke victims might 'write' letters using the application to replicate their own handwriting. Remote signing of greeting cards. etc. It's believed the technology can also be used to identify fake writings, and signatures. It is said it could also be used to translate handwriting into other languages, and still retain the handwriting of the original author. ~ Input data required to produce credible text is as little as four or five words. However, it's not clear what was used in developing the Machine Learning algorithms in the links I've reviewed. ~ The machine learning algorithm is based on 'glyphs' which is any given instance of a character. Each individual has a style in their handwriting that can be recognised in the glyph samples. The software learns the consistencies within the style of that author and uses these to write. Pen lines, colours and textures are learned, joins between characters are used, as is spacing.  ~Individuals asked to decide whether text was written by humans or software were fooled 40% of the time. The image shown of the handwriting samples on the link above looks very credible.  ~Concern was expressed in the use of such software in the forging of documents. The team counter that their software could be used to detect forgeries by using 'texture synthesis'. Further links: http://visual.cs.ucl.ac.uk/pubs/handwriting/ 
72 ~191 ~Nvidia Research: Production-Level Facial Performance Capture Using Deep Convolutional Neural Networks (http://research.nvidia.com/publication/facial-performance-capture-deep-neural-networks) ~Samuli Laine, Tero Karras, Timo Aila, Antti Herva, Shunsuke Saito, Ronald Yu, Hao Li, Jaakko Lehtinen ~Generate depth facial modeling from video capture. ~Videos from nine different camera. ~CNNs, data augmentation and GPU acceleration. ~Compare to other traditionnal method (Thies, Cao), the recognition of facial expression is way more performant. The Nvidia system can reproduce more complex facial expression. ~
117 ~325 ~https://www.devex.com/news/bringing-machine-learning-to-last-mile-health-challenges-91453 http://www.selectscience.net/SelectScience-TV/Videos/easyscan-go-optically-intelligent-disease-scanning?videoID=4022 https://spectrum.ieee.org/the-human-os/robotics/artificial-intelligence/aipowered-microscope-counts-malaria-parasites ~The Global Good Fund and Motic (a China-based company that specialises in manufacturing microscopes). ~To automate the analysis of blood smears in the testing for malaria. This will provide faster turnaround time for patients, less time and effort required by lab technicians, and data for tracking and analysing the spread of different strains of the parasite. ~Thousands of slides of blood samples from potential malaria patients. ~Computer vision algorithms and deep learning. ~An experimental version of the microscope has already shown that it can detect malaria parasites well enough to meet the highest World Health Organization (WHO) microscopy standard (competence level 1). This means it is on par with well trained microscopists. ~No comment. 
148 ~211 ~Forbes.com "LexSet Is About To Revolutionize Interior Design With NextGen AI" https://www.forbes.com/sites/westernbonime/2017/11/19/lexset-is-about-to-revolutionize-interior-design-with-nextgen-ai/#5d5e7afb7976 ~Western Bonime, contributer to Forbes #NewTech ~LexSet is a family of applications which aim to use machine learning to provide people with  smarter Interior Design suggestions. ~A large data-set of professionally curated interiors. Images of furniture available at a number of retailers. Images which the user provides of their own interiors. METHOD LexSet compares images of the users interior with those of the professionally curated interiors. It also applies a number of algorithms specific to the principles of good design. ~LexSet returns a set of more suitable furniture suggestions to the user as well as information on where to purchase certain furniture. ~Article is written as though it was a paid advertisement. Very biased towards the success of the application and how it will "revolutionize the way people shop, assemble and design for their homes" 
191 ~123 ~The Conversation https://theconversation.com/googles-new-go-playing-ai-learns-fast-and-even-thrashed-its-former-self-85979 ~Google/AlphaGo Zero ~To improve on the original AlphaGO program. ~The data that the program was given was just the rules of the game. ~The program used a combination of two of the most powerful algorithms currently fuelling AI: deep learning and reinforcement learning. AlphaGo Zero never saw humans play. From a relatively modest five million games of self-play, taking only three days on a smaller computer than the original AlphaGo, it then learned super-AlphaGo performance. Its learning roughly mimicked some of the stages through which humans progress as they master Go. AlphaGo Zero rapidly learned to reject naively short-term goals and developed more strategic thinking, generating many of the patterns of moves often used by top-level human experts. But remarkably it then started rejecting some of these patterns in favour of new strategies never seen before in human play. ~The AlphaGo Zero program made a huge improvement, AlphaGo Zero beat the original AlphaGo in a 100-game match by 100 to 0. ~It was very impressive that no humans were involved in the ALphaGo Zero training, it poses the question, will computers need humans at all in the future? 
211 ~239 ~Techdirt.com "Why Netflix Never Implemented The Algorithm That Won The Netflix $1 Million Challenge" https://www.techdirt.com/articles/20120409/03412518422/why-netflix-never-implemented-algorithm-that-won-netflix-1-million-challenge.shtml ~Netflix.com ~The goal of Netflix was to improve their current success rate in predicting user ratings of certain films by 10%. This would improve the way that they recommended new content to users. The competition was open to the public to come up with such an algorithm with 1 million dollars offered to the first group to do so.  ~Flat files containing demographic information about users and the ratings they had given certain films. Information was given in a separate flat file about these films such as their genre and release year.  ~The system that Netflix eventually implemented for recommendations by Netflix was a mixture of Restricted Boltzmann Machines and Singular Value Decomposition ~Singular Value Decomposition produced a root mean squared error of 0.8914, Restricted Boltzmann Machines produced a root mean squared error of 0.8990. A blend of both models reduced this error to 0.88.  ~Interestingly, this was not the winner of the challenge as it didnt improve the Netflix algorithm by 10%. Netflix found that the winning entry which did so was too challenging to implement and that the small advantage it gave over simpler models wasnt enough to justify the investment.  
120 ~0 ~~ Apple   ~  To provide a voice for Siri that is both pleasant and intelligible, which covers all possible utterances that siri may speak. ~ Between ten to twenty hours of speech patterns. These patterns can range from readings from a novel to jokes from a joke book.  ~ These ten to twenty hours of speech is not enough to cover all possible speech patterns siri may need, thus apple slices the recorded speech into units. Apple then uses a deep mixture density network to model both the means and variances of the speech features which is used for guiding unit selection sythesis. ~A deep MDN-based hybrid unitselection text to speech system for siri. ~ NULL 
146 ~116 ~https://techcrunch.com/2017/11/16/algorithmia-now-helps-businesses-manage-and-deploy-their-machine-learning-models/ ~Algorithmia ~To launch a new online platform which hosts and runs machine learning algorithms while allowing them to be shared with colleagues and analysed. ~Algorithmia started as an online marketplace for algorithms. As such, they have an immense catalogue of knowledge on algorithms, including machine learning algorithms, and have extensive infrastructure for the hosting of these algorithms. The data they have accumulated allows them to know about use-cases for algorithms, computational costs, complexity etc. ~Algorithmia has deployed a suite of tools which allow data scientists to create machine learning algorithms in any language and using any framework that they please. These are then hosted in the Algorithmia cloud (CODEX platform).  ~The system is already being tested and used by U.S government agencies including the CIA and large organisations who will rely on the scalability, reliability, security and shareability of the new platform. ~ N/A 
184 ~90 ~https://www.devex.com/news/bringing-machine-learning-to-last-mile-health-challenges-91453 ~ Global Good Fund - joint venture between Bill Gates and Motec, a China based company  specialising in microscope manufacturing. ~To identify and count the number of malaria parasites in a blood sample to help and fight the disease. ~Blood samples taken from those who are known to be infected with malaria - there are  millions of cases every year so there is endless data ~The microscope, called the EasyScan Go, will be installed with image recognition software so that it can 'see' the blood smears and be trained to identify the parasites. It will use a machine learning algorithm to get better at identifying the parasites. ~At present the researchers have shown the the EasyScan Go is as reliable as an expert microscopist in identifying the parasite from a sample. ~The possibilites for a technology like this are endless - it could lead to any disease which is held in the blood being able to be identified by machines instead of humans. This would mean that more cases would be caught early on and so would lead to less deaths. 
74 ~72 ~ apple https://machinelearning.apple.com/2017/11/16/face-detection.html ~ Apple ~to use Machine learning algorithm and deep learning to detect and identify individual faces    ~ pictures of peoples faces  ~ OverFeat drew the equivalence between fully connected layers of a neural network and convolutional layers with valid convolutions of filters of the same spatial dimensions as the input ~ you can login to your phone via facial recognition  ~lot more detail on the site  
167 ~78 ~https://gmail.googleblog.com/2015/07/the-mail-you-want-not-spam-you-dont.html ~Google ~To automatically detect and remove spam emails. ~Assessed emails that were frequently marked as spam by Gmail users, looking for common words or phrashes that would show that it was spam. ~Since email spammers are constanly updating their spam to bypass the latest in blocks and filters, the google spam detector also constantly updates itself, in an attempt to counteract the spammers. It constantly learns, as each time a user reports an email as spam, the AI can assess it and see how it can detect that kind of email in the future. ~Google's spam filter has produced some excellent results. By their own esimate, less than 0.1% of email in the average Gmail inbox is spam, and the amount of wanted mail landing in the spam folder is even lower, at under 0.05%. ~This is a service that affects me first hand, as a Gmail user. I can't remember ever having a problem with receiving spam emails, or having wanted emails marked as spam. 
228 ~239 ~https://thenextweb.com/artificial-intelligence/2017/11/13/alibabas-ai-is-the-blueprint-for-brick-and-mortar-stores-of-the-future/ https://www.technologyreview.com/s/609452/alibabas-ai-fashion-consultant-helps-achieve-record-setting-sales/ ~ Alibaba reserchers and developers ~To make smarter shopping sugegstions for offline customers in stores, using machine learning to suggest items that the customer usually looks at and buys with another item. They aim to make shopping at physical stores as seamless and as efficient as shopping online. They aim to stop the decline in offline shopping. ~Customer's known shopping data such as items they usually put in their cart together (for example hats with belts, shoes with socks, etc) and what they end up buying when they checkout. ~They used a deep learning algoriths to allow the AI to make smarter shopping decisions for the customer. Sensors were placed in each garment of clothing in the store so that once in the changing room a large screen could allow the customers to try on the garment in VR along with similar suggested garments. ~The combination of this new technology known as FashionAI and ALibaba's existing technology allowed them to sell  25 billion dollars worth of goods in one day. ~Clearly the FashionAI was a success in the 13 stores it was installed in for the trial run. It gives people an incentive to physically shop instore and if it is extended to widespread use across the world in major retailers it could stop the decline in brick-and-mortar shopping and benefit retailer and consumer alike. 
108 ~100 ~https://venturebeat.com/2017/11/09/voice-recognition-and-machine-learning-make-service-bots-better/ ~Peter Quinlan, Tata Communications ~Utilizing advancements in AI to implement 'robots with customer service skills' ~The bot speaks to the customer in an accent they will understand based on their accent. This requires large datasets of accents. Similarly, the system knows all the customers past experience/communications with the company. ~The bot addresses the customers concerns using the collected wisdom of the company, including the absolute latest data on issues other customers face and how to solve them. Can provide solutions without errors. Follows up to make sure solution worked. ~Potential to increase customer satisfaction and loyalty as well as making the customer service experience more seamless. ~None. 
73 ~40 ~http://ori.ox.ac.uk/how-robotcar-works/ ~ORI Autonomous Systems ~Creation of 'infrastructure-free' navigation, in particular in self-driving cars. ~Data is collected from two lasers tucked under the front and rear bumpers of the car, which allow the car to generate a 3D structure of the world around it. ~The specific techniques used are not discussed at all. ~Results are not discussed, however their description of the system heavily indicates that it is still very much a work-in-progress. ~N/A. 
126 ~64 ~Graphical Models for Machine Learning and Digital Communication ~Brendan J. Frey ~To Solve the problems in machine learning and digital communication deal with complex but sutured natural or artificial systems. ~HE used Graphical models as an overarching framework to describe and solve problems in the area of pattern classification, unsupervised learning, data compression and channel coding. ~Bayesian belief networks and Markov random fields. ~Error correcting codes used in the telephone modems and deep space communication consists of electrical signals that are linked together in a complex fashion determined by the designed code and the physical nature of the communication channel. ~With great clarity the cutting edge of research on the learning of graphical models, the compression of data using latent variable models, and the channel coding. 
316 ~149 ~CINNYINNO website, link: https://www.americaninno.com/cincy/cincy-startups/is-that-place-a-hipster-jungle-spatial-uses-machine-learning-to-find-out/ ~Spatial, a Cincinnati startup that work at an API for behavioural location data. The products they offer fall under 2 branches, the neighbourhood API and the POI (Point of Interest) API. The neighbourhood API reveals the current behavioural trends of any specified city, neighbourhood, or custom area. They also integrate their pre-built tool to overlay neighbourhood personalities on your existing map. POI data can access unique behavioural profiles of neighbourhoods that provide a better understanding of POIs to any existing dataset, as well as delivering relevant POIs of for customers' "natural language" queries. In short, with ethnography and machine learning, they created an AI that understands human behaviour around locations. Link: http://spatial.ai ~Wanted to get a better indication as to activities that occur around areas/locations based on their behaviours and personalities. ~They aggregated data from about 30 sources which included social media sites such as Twitter and Instagram. METHOD They then applied Machine Learning models to possibly clustering models that are used to get a strong indication on different behaviours in any area provide, that also follows a "social point" score system, the closer a data point is to a highly suspected behaviour, the higher the score will be. ~They managed to create an API that gives people information about areas in terms of behaviour or personality, whether the area is "artsy" or even "vegan". Ultimately adding a new informational layer of areas that people search up and are curious about on a behavioural/cultural level. ~I find this truly interesting because it adds to something that was never thought about beforehand. It enhances the experiences of learning about different areas, breaking out of the traditional mapping tools. Instead of just asking for a neighbourhood, you can ask the search engine if it is popular with families in certain regions, once again giving a new dynamic to map tools as we know it.    
80 ~260 ~http://uk.businessinsider.com/visualdx-machine-learning-app-for-skin-diagnosis-ceo-interview-2017-11?r=US&IR=T/#visualdx-is-intended-for-use-by-doctors-to-confirm-and-validate-diagnoses-it-allows-doctors-to-search-by-symptoms-signs-and-other-patient-factors-1 ~VisualDX an app for doctors ~to use the data to help indentify diferent skin conditions in particular. identify disease or condition depicted in photos. ~built in database of 32000 high quality medical images. METHOD Used coreML as software that can run machine learning algorithms on the phone. RESULT ability for doctors to look up sympthums and get a better idea of what could be wrong. Further development in an app for the end users for just skin conditions. ~NA 
49 ~86 ~http://www.mining-technology.com/features/ever-learning-mineral-exploration-platform/ ~Ph.D. student Roman Teslyuk, Sydney University ~Predict the mineral composition of areas to help companies find mineral reserves. ~Geophysical, satellite, geochemical and user uploaded data. ~Unsupervised machine learning to create a data-driven geological map. ~Geological map of predicted mineral deposits. ~Vague on quality of results and algorithms used. 
99 ~82 ~ http://news.mit.edu/2016/eye-tracking-system-uses-ordinary-cellphone-camera-0616 ~ Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory and the University of Georgia ~to  where mobile-device users are looking.  ~ training set includes examples of eye-gaze patterns from 1,500 mobile-device users. ~ they use computer vision to track the eyemoment of the user and used the data set to train the machine to determine the user eye ~ they were able to track where eye is looking by using simple front camera of most common used smartphone ~this could really be helpful in determining how users react to advertisemnets and placement of it 
203 ~251 ~https://qbeeurope.com/news-and-events/press-releases/england-and-all-blacks-to-win-all-autumn-internationals-games/ http://www.bbc.com/sport/rugby-union/international-match/results https://blog.passle.net/post/102dz5e/rugby-content-marketing-from-qbe ~QBE Business Insurrance. ~Predicting result of rugby matches. ~Inolves a range of data including points scored, caps of each captain, world ranking and match location. ~Methods were not disclosed. (Complex Mathmatical formula.) ~-Predicted- Scotland 31 Samoa 22 England 38  Argentina 21 Wales    20  Australia 24 Ireland   20  South Africa 20 France  16   New Zealand 35 Italy       20   Fiji 27 -Real results- Scotland 44 Samoa 38 - Result: Correct, Margin: Over by 3 England 21  Argentina 8 - Result: Correct, Margin: Over by 4 Wales    21  Australia 29 - Result: Correct, Margin: Under by 4 Ireland   38  South Africa 3 - Result: Incorrect, Margin: Under by 35 France  18   New Zealand 38 - Result: Correct, Margin: Under by 1 Italy       19   Fiji 10 - Result: Incorrect, Margin: Under by 2 -Accuracy: 66% ~The formula had also been used to correctly predict last years six nations champions. It has an 80% overall success rate and has come close to exact margins/scores over the years. 
440 ~111 ~Written by Charlotte Hu https://www.teslarati.com/tesla-autopilot-ai-artificial-intelligence-unfair-advantage/ ~The agent here is Tesla and its autopilot used in Self driving cars. ~The goal is to have an autopilot that is reliable. A big factor in the way of that is the fact that machines have no moral compass. How can they make snap judgements on what is the best decision in the face of an accident.  ~There is much data used from all areas of driving. From people and objects on the road, gps data, car data like speed etc, data on the weather conditions etc. As tesla has a network of connected cars on the road they are gathering data all the time and looking for ways to improve autopilot with that data and methods in machine learning.  ~Historically Tesla focused on "narrow AI". Narrow AI enables the car to make decisions without driver interference. "The vehicles would increasingly rely on radar as well as ultrasonic technology for sensing and data-gathering to form the basis for Tesla's Autopilot algorithms. A technology that isn't derived from LiDAR, the combination of radar and camera system said to outperform LiDAR especially in adverse weather conditions such as fog". "With the introduction of Autopilot 2.0 and Tesla's 'Vision' system, and billions of miles real-world driving data collected by Model S and Model X drivers, Tesla continues to create a detailed 3D map of the world that has increasingly finer resolution as more vehicles are purchased, delivered and placed onto roadways. The addition of GPS allows Tesla to put together a visual driving map for AI vehicles to follow, paving the path for newer and more advanced vehicles." Apart from the main algorithms that are controlling it I am unaware. They have taken on board Andrej Karpathy as the new Director of Aritificial intelligence at Tesla. That could be considered a method for bringing autopilot to the next level. Karpathy came from Open-AI, Elon Musks non profit AI company.  ~Tesla's autopilot is already capable of self driving and it is only to get better with the use of machine learning. Using ML with their masses of data will greatly improve the performance of the cars autopilot. I would already considered it a success  but the fact still remains on the car having the ability to make moral judgements in the right dangerous situations.  ~Although the autopilot is already at a very advanced stage, it will take a considerable amount of work to get it to a stage were it can make consistant moral judgements in dangerous scenarios, but I fell in it is in the right hands with Tesla and Elon Musk.  
22 ~117 ~https://www.computerworld.com/article/2860758/ibm-detects-skin-cancer-more-quickly-with-visual-machine-learning.html ~IBM ~Detect skin cancer more quickly ~3000 scan images ~Cognitive computing based visual analytics ~95 percent accuracy in spotting melonoma ~
94 ~106 ~https://www.newscientist.com/article/2151268-an-ai-has-learned-how-to-pick-a-single-voice-out-of-a-crowd/ ~Niels Meinke ~To solve the noisy coctail problem.It means identify a specific voice among a crowd. ~The algorithm trained using 100 English speakers. METHOD They used deep clustering to identify matchles voiceprint of 100 English speakers. ~The system is now able to differentiate two people speaking into same microphone with 90 percent accuracy. If the number of speaker is increasing, the percent is still up to 80. ~It is pretty amazing to see this kind of success rate, if you consider even humans can have a hard times about classifying voices and faces. 
155 ~64 ~http://news.mit.edu/2017/bug-repair-system-learns-example-0928	 ~Martin Rinard, Fan Long and Peter Amidon of University of California, San Diego. ~To create a program that can automate the process of code-patching with the use of templates that indicate the general form that patches tend to take. ~Dataset consisting of original, buggy code and the patch that was applied to it. ~A draft template is generated from every possible pair of training examples. The examples that generate the most promising templates are augmented with each of the other examples in turn to create a large set of three-example training sets. New draft templates are generated for these three-example sets and the whole process is repeated a total of four times to get a set of excellent final templates. ~The code-patching system was able to patch nearly twice as many bugs as previous systems that didn't utilise machine learning. ~The system could possibly be further improved with the help of a bigger dataset. 
353 ~323 ~Gizmodo.com - Youre Using Artificial Neural Networks Every Day Online  Heres How They Work https://gizmodo.com/youre-using-neural-networks-every-day-online-heres-h-1711616296 Google  The Neural Networks Behind Googles Voice Transcription https://research.googleblog.com/2015/08/the-neural-networks-behind-google-voice.html ~Google ~Voice Transcription  The implementation of a system that determines with near perfect accuracy what a person is saying in a variety of languages. ~A huge training set of audio recordings accompanied by scripts of what exactly is being said. A great source of this for Google were things like movie and tv show scripts and the audio from them. This provided them with a near unlimited training set to work on. ~The method employed for this task has evolved over time, starting with a Gaussian Mixture Model and moving onto a multi-layered Deep Neural Network combined with back-propagation. The first layer in the network chops up an audio and identifies vowel sounds. The subsequent layers work to figure out how vowel sounds might be combined to form words. Whenever the neural network misidentifies a word the system back-propagates, adjusting the weights throughout the inner layers of the neural network in order for it to correctly identify the audio on the next occasion. Google then moved onto using Recurrent Neural Networks for the task, which in addition to the functionality of a Deep Neural Network have the ability to remember what they have heard so far, much like humans recognise speech based largely on what has already been said.   ~When Google first released their Voice Transcription onto the public market in 2011, the error rate was around 25%. Over the next six years changes to the types of neural networks used as well as more data being available to Googles models has led to that error rate being reduced to under 8%. ~One of Googles greatest challenges was correct punctuation. Punctuation and how it is used varies greatly across languages and very subtle differences in speech emphasis can completely change the meaning of speech and with it the correct punctuation. A big breakthrough came with Googles use of Recurrent Neural Networks which gave Googles models context, greatly assisting in the determination of correct pronunciation.  
105 ~99 ~Example 6: Understanding deep features with computer-generated imagery https://www.robots.ox.ac.uk/vgg/publications/2015/Pfister15a/pfister15a.pdf ~Tomas Pfister   James Charles Andrew Zisserman  ~human pose estimation in videos, where multiple frames are available  ~BBC Pose dataset,Extended BBC Pose dataset, ChaLearn dataset and Poses in the Wild (PiW) and FLIC datasets. ~Compares the estimated joints against frames with manual ground truth ~Created a new architecture for pose estimation in video which could be expanded further to things such as classification and segmentation ~Another bedrock project which will be used further to improve neural networks ability to render 3d scenes and allow future CNN's in the to produce better 3d rendering. 
324 ~100 ~ https://support.apple.com/en-us/HT208108 https://9to5mac.com/2017/09/13/face-id-demo-fail-details/ ~ Face ID is developed by Apple. It claims to be a more intuitive and secure way to authenticate a user. ~ Face ID's goal is to provide a secure means of unlocking your iPhone. It can also be used for purchasing items from iTunes Store, App Store, iBooks Store and for payments with Apple Pay, and to sign in to applications. ~ The application uses learning algorithms to learn the users features over time. "The TrueDepth camera captures accurate face data by projecting and analyzing over 30,000 invisible dots to create a depth map of your face and also captures an infrared image of your face. A portion of the A11 Bionic chip's neural engine — protected within the Secure Enclave — transforms the depth map and infrared image into a mathematical representation and compares that representation to the enrolled facial data." ~ The authentication method reacts to changes in appearances. If a change is too great there is an option for authentication by using a passcode before updating changes to the users appearance.  The application is triggered when the iPhone screen is tapped, when a notification wakes the screen, or when the iPhone is raised. There are additional security measures including extra validation if the iPhone has not been unlocked in 24 hours and if the iPhone in question has been unlocked remotely.  ~ During the first public demonstration of Face ID there was a slight mishap; the iPhone would not unlock using face recognition. The iPhone in question had been handled by a multiple of people before setting up for a live demonstration by Craig Ferderighi.  ~ Apple claim that the “Face ID worked as it was designed to” as it required a passcode to unlock. What worries me is that it was possible for the faces of others to influence this at all (and apparently without their knowledge during their handling of the iPhone). 
227 ~271 ~ http://fortune.com/2017/03/28/amazon-go-cashier-free-store/ https://www.amazon.com/b?node=16008589011 https://www.quora.com/How-is-deep-learning-implemented-in-Amazon-Go-special-grocery-stores-in-which-you-pick-up-whatever-you-want-and-then-leave-without-waiting-to-pay ~ Amazon Go aims to provide a convenience store that epitomises convenience. Amazon's team began developing this system four year ago. At the moment the store in in Beta and is only available to Amazon employees. ~ Amazon Go is a checkout-free shop that allows its customers to simply walk in, acquire what they need, and leave without having to queue to pay. In order to participate in this shopping experience, first the customer must create an Amazon account and download the Amazon Go app. ~ Amazon uses SnapTell (bought company in 2009) integrated with other technologies. Amazon claim that the same technologies used for self-driving cars has been used for Amazon Go. Over time the behaviours of customers are tracked, and will be used to improve future shopping experiences. ~ Computer vision, sensor fusion and deep learning all combine to track users and items alike.  ~ There have been some reported issues surrounding tracking many customers during one time. The sensors are unable to successfully track more than 20 people approx. ~ It is not a question of "if" the store will be released to the public, but "when". There may be more issues in to future surrounding people trying to dupe the cameras and sensors, but I predict that in the future the amount of successful shoplifters will steadily decline. 
208 ~193 ~Journal of the American College of Radiology  http://www.radiologybusiness.com/topics/technology-management/just-beginning-6-applications-machine-learning-radiology-beyond-image-interpretation ~Paras Lakhani, MD, department of radiology at Thomas Jefferson University Hospital in Philadelphia, and colleagues. ~To help solve problems in radiology image interpetation. Including the likes of being able to reduce radiation does during CT scans which would, with Machine Learning, make it safer and help produce better images given this drop in dosage. It can be used to reduce the time it takes to take an MRI scan in the same way, being constucted from raw data from the scanner meaning there could be a possible  drop in 50% in time for scans. With better knowledge of images and the data they contain in a medical sense, The possibility of an image search engine couold vastly improve the "guess work" with images that are difficult for humans to clasify. ~electronic health care databases including images in PACS, radiology reports and ordering information in Radiology Information Systems. electronic health records that encompass information from other sources, including clinical notes laborator data and pathology records. radiology images rich in metadata stored in the DICOM format. METHOD Deep Learning. ~None Mentioned. ~This is a relatively new article and is undergoing. can help interpet reports more accurately. (doc's got bad handwriting) 
102 ~59 ~ http://fortune.com/2015/10/16/how-tesla-autopilot-learns/ ~ Tesla ~ Self driving cars capable of out performing human pilots and reducing accidents on the road. ~ Thousands of hours of driving footage, from dash-cams to traffic cameras. ~  The algorithms use visual techniques to break down the videos and to understand them. The goal is that when something unexpected happens - a ball rolls into the street - the car can recognize the pattern and react accordingly. ~ Hopefully a self driving smart car capable of avoiding collisions. ~ Self driving cars are still in the early stages of development however tests are showing positive signs. 
155 ~141 ~www.newscientist.com/article/2146703-even-a-mask-wont-hide-you-from-the-latest-face-recognition-tech/ (https://arxiv.org/pdf/1708.09317.pdf) ~Amarjot Singh, Devendra Patil, G Meghana Reddy, SN Omkar ~To enable face detection technologies to recognise faces even if the subject wears a disguise. ~Two face disguise datasets with 2000 images each. The databases are composed of images of 25 subjects ranging in age from 18 to 30 and with a variety of backgrounds (simple and complex), and 10 different  disguises.   ~14 key points on the face are used to create star-net structures to classify images. Pictures of the disguised face is then compared against a group of 5 individuals without disguises, including the disguised subject, and the similarity of a disguised face is estimated against the non-disguised face by computing an L1 norm between the orientation of different key points obtained using the net-structure ~The frame-work is shown to outperform the state-of-the-art methods on key-point detection and face disguise classification ~Will be useful for law enforcement and intelligence communities. 
237 ~128 ~Hyeokjun Choe, Seil Lee, Hyunha Nam, Seongsik Park, Seijoon Kim, Eui-Young Chung, Sungroh Yoon https://arxiv.org/abs/1610.02273 ~Cornell University Library. research and teaching for Cornell university. ~During algorithm training in ML alot of big data transfer takes place which demand a great deal of computation.the goal is to introduce methods to solve this problem. ~The possible improvements of making further advancements in accelerating computing in this paper was evaluated for machine learning using a new platform they have developed, based on machine learning workloads. ~SGD- Stochastic  gradient descent is used for training differentiatable models such as neural networks  NDP-Near Data Processing this is the simple idea of placing the processing power near the data rather than shipping the data to the processor. NDP is motivated by the cost of data movement. ISP-Instorage Processing similar to NDP ISP tries to minimise the movement of data by running applications on processors in the storage controller.ISP is generaly used by small companies.  using ISP we compare different SGD variants and compare the performance of ISP over conventional in-host processing method. ~The results of the experiment showed that ISp provides a solution that can potentially reduce the issues with data transfer by processing core operations in the storage level.compared to the in host processing which slowed down inaccordance to reduced memory size. The NDP implementation reduces the number of data transfers and offloads some computation burden of the CPU. ~
107 ~165 ~Online article on The Guardian https://www.theguardian.com/technology/2017/nov/09/self-driving-bus-crashes-two-hours-after-las-vegas-launch-truck-autonomous-vehicle ~Navya, a French company. ~To build and deploy a successful self-driving bus in Las Vegas, Nevada, USA. ~Imaging data necessary for implementing this project. ~Mathematical analysis, machine learning, neural networks, and computer vision techniques. ~The trial was going well until it hit a vehicle about two hours after deployment. No reported injuries were noted. The bus did not have the ability to reverse and thought that stopping would prevent an accident with the truck. ~I have always been skeptical of self-driving vehicles. The technology will never be perfect. It was good to know that it was not a huge accident nevertheless. 
125 ~109 ~http://www.digitaljournal.com/business/how-machine-learning-is-shaping-up-investing-interview/article/507704 ~A tool called Market Sensei by a company called Expat Inc. Ran by founder Patrick Kwete ~use patented machine learning algorithms to give powerful prediction abilities to both experienced traders and novice investors ~The data has been used from how the stock market has performed in the past and is updated daily. It uses this  to continously update its metrics in real time to allow it to pass the user a prediction for  the next 7 days. ~Patented Algorithms built inhouse without Frameworks ~Boasts a 60-80% accuracy in stock fluxuation within the previous 9 days.  ~managed 6000+ investors and paying investors from the first month. have a intuition game built into it and insights to help even successful traders be better.   
77 ~78 ~https://www.macrumors.com/2017/11/16/machine-learning-journal-face-detection/ ~Apple. Published on Apple's machine learning journal ~To EFFICIENTLY detect similar faces in images ~There are two sets of images. An entire library of images, and the individuals personal images ~Standard neural nets are used for facial recognition, but forms of cluster sampling are used to group these faces. Apple write low-level algorithms to utilize their iPhones CPU and GPU for optimum efficiency ~Succesfully groups faces from images with high percentage success. Improving every year ~None 
105 ~106 ~https://www.technologyreview.com/s/609469/this-ai-learns-your-fashion-sense-and-invents-your-next-outfit/ ~Wang-Cheng Kang, UC San Diego Chen Fang, Adobe Research Zhaowen Wang, Adobe Research Julian McAuley, UC San Diego ~Use generative adversarial networks (GANs) to learn and predict a persons preferred style of clothing. ~Data scraped from Amazon: shoes, tops, and pants (all for both women and men). ~Convolutional neural networks (CNNs) were used to learn and classify a users preferences, which was then used to train a GAN that could generate fake images of clothing that the user may like. ~While good at generating specific items of clothing matching the persons style, it has no real sense of what makes a cohesive outfit. ~N/A. 
317 ~199 ~Medium Corporation, The Physics arXiv blog; https://medium.com/the-physics-arxiv-blog/when-a-machine-learning-algorithm-studied-fine-art-paintings-it-saw-things-art-historians-had-never-b8e4e7bf7d3e ~Rutger University. Researchers:  Babak Saleh, Ahmed Elgammal Rutgers University in New Jersey ~Classifying fine art normally involves finding out who the artist is, the style used, the genre, age of the painting and recognising influences between artists.  This can be a difficult task, identifying influences can be particularly difficult and usually requires art experts to do the work.  The goal was to create an algorithm that would carry out the task of classifying paintings and identifying influences between artists ~The data they used came from a large database of digitised images of paintings by more than a thousand artists.  The images had been catagorised according to styles and genre.  For their training set they used over 1700 paintings ranging from the 15th to the 20th century and painted by 66 different artists in 13 different styles.  ~The algorithm used machine learning and machine vision.   The images of the paintings where broken down into over 2500 classemes.  Classemes are classifiers describing features such as colour, form, shape, type of object (eg a swan, swimmer, body of water, etc) The result was a vector like list of words describing the painting.  The paintings were then compared by looking for similar vectors ~Opinions from experts on paintings and artists who had be influenced by others were gathered and used to measure their results.  The algorithm used was able to correctly identify the artist in 60% of cases and the style in 45%.  The algorithm was also found to be able to identify paintings that had been influenced by other paintings.  Using the algorithm, they discovered influences between paintings that had not previously been recognised by art historians.  ~The researchers say that a refinement of the same method could also be used for other types art e.g. music or literature. 
127 ~176 ~http://markets.businessinsider.com/news/stocks/Market-Sensei-Launches-Machine-Learning-Powered-Stock-Market-Prediction-Platform-for-Novice-and-Experienced-Investors-1002298701 ~Expat Inc. ~To deliver an app capable of predicting the rise and fall of thousands of shares on the stock market. ~Data is taken from previous stock price changes on NASDAQ and DOW along with vast amounts of financial and individual company data. ~Expat's Market Sensei algorithm is patented but it is safe to guess they use some variant of one of Autoregressive integrated moving average , Kalman filter or Recurrent neural network. RESULT A transparent analytics platform that provides accurate predictions to individual stock price rises and falls. ~A useful tool for those who want to get into trading on the stock markets. Market Sensei is transparent as it lets the user know its prediction rate allowing users to make a well informed decision pre-trade. 
136 ~110 ~https://www.newscientist.com/article/2146703-even-a-mask-wont-hide-you-from-the-latest-face-recognition-tech/ ~Matt Reynolds ~To create a facial recognition software to identify a target who is hiding their face (partially or completely)  ~The system was tested on a handful of applicants, wearing a range of clothing to obscure facial features. ~The system was rated on a success to fail ratio. ~System identified someone wearing a scarf 77% of the time, a cap & scarf 69%, scarf and glasses 55% of the time. ~Although the system is not as good as one to recognise people without clothing to obscure facial features it is interesting to see there is technology that can work around this obvious downfall of facial recognition. In particularly this could help police officers, who could find great use of software that recognises a known perp. Regardless of them having clothing to obscure their face. 
108 ~114 ~http://www.cancernetwork.com/articles/computer-technology-helps-radiologists-spot-overlooked-small-breast-cancers ~Doctors ~Many lesions in early stage breast cancer are near invisible and others might not be detected by tired or less experienced radiologists, CAD intends to improve the detection rate. ~Over 22,000 women were screened routinely over 5 years and their mammograms were reviewed by a CAD prototype. ~A laser scanner transforms the mammography film into a matrix of data. Microcalcifications are shown as white spots and masses as round or irregular shapes. The system uses computer vision and artificial intelligence algorithms to sift out soft tissue and highlight patterns indicating lesions. ~The CAD identified 52% of these missed cancers a year before they were detected. ~
101 ~112 ~https://techcrunch.com/2017/11/08/skype-launches-photo-effects-sticker-suggestions-powered-by-machine-learning/ ~ Microsoft  ~Automatically add photo stickers to photos based on the content and/or date. ~All of the photos taken on microsoft devices. ~The machine learning algorithm likely bases the photo sticker off the content of the photo which is conpared to it’s training set and then they add a weighting factor for the days of the week. ~The results of this is a system to automically modify the end users photos. ~No details about the method is given in the article so the METHOD section is purely speculative. Also the system is likely heaily weighted towards the seasonal/daily stickers. 
93 ~75 ~https://www.technologyreview.com/s/406637/the-1-million-netflix-challenge/ ~Netflix ~To make recommendations more accurate ~100 million recommendations which were striped of user date ~Methods varied but the common core was finding pairs of users who gave shows the same rating and then recommending shows based off that.  Small variations included giving different weighting based off the day of the week a show was watched, how many show were rated at one time and the length of time between a show being watched and rated. ~They were able to improve the success rate of the recommendations system by 10% ~
298 ~86 ~https://arnesund.com/2015/05/31/using-amazon-machine-learning-to-predict-the-weather/ ~Amazon Machine Learning Service ~To use Amazon Machine Learning (AML) to Predict the Weather ~The dataset chosen contains weather observations from five cities in Norway, scattered around the southern half of the country. The weather in Oslo usually comes from the west, so observsations from cities like Stavanger and Bergen were included in the dataset also. To create a dataset with enough data in each record to be able to predict the target value, all weather observations were appended with the same timestamp, regardless of location, to the same record. This means that for any given timestamp there will for instance be five temperatures, five wind measurements and so on ~Binary classification, Multiclass classification, Regression, The study was tested by trying to predict the temperture at a certain time, in a certain city in Norway, in this case Oslo, AML uses in built algorithms that can be easily selected, this allowed the study to be done ~Taking into account the positive bias of 1 to 2 degrees and a prediction mean value of 13.6 degrees Celsius, the actual temperture was 12.0 degress Celsius, To improve the model performance further the bias could be reduced further. To do that the model would have to be trained with more training data, since the dataset was small on this occasion. To get a model which could be used all year around, the training data from a relevant subset of days throughout the year (cold days, snowy days, heavy rain, hot summer days and so on) would have to be included. ~The dataset was small so this study was not incredibly accurate, but I think it could be useul if an extensive study was performed, it could be easy to test using historical data. however the weather is hard to predict. 
108 ~112 ~https://www.technologynetworks.com/tn/news/scientists-use-machine-learning-to-analyze-language-in-movies-294179 ~ Researchers at the University of Washington. ~ To analyse the significance of a character in a movie based upon their scripted dialogue. ~ They used data obtained from the scripts of nearly 800 movies. ~ They used natural language processing tools, which rely on machine learning algorithms to identify characters that appeared as a verb's subject and object. ~ The results showed a tendency for male characters to score higher on the tests of significance than female characters. ~ Implications of this can be used by future writers to see how much significance they are giving to a character based on the dialogue they are writing. 
87 ~137 ~https://thenextweb.com/artificial-intelligence/2017/10/20/googles-deepmind-achieves-machine-learning-breakthroughs-at-a-terrifying-pace/ ~Google Engineers ~To create an ever smarter AI that can learn without any human lessons. ~Not Available ~Googles DeepMind ~Alpha Go Zero has defeated Googles previous version Alpha Go(The computer that beat the worlds greatest human go player) without a single human lesson. ~Alpha Go Zero has only four AI processors and the only data it was given was the rules of the game. Nobody taught it how to play or fed it thousands of matches to study, and yet it outsmarted the original Alpha Go. 
159 ~116 ~https://cloud.google.com/blog/big-data/2016/08/how-a-japanese-cucumber-farmer-is-using-deep-learning-and-tensorflow ~Makoto Koike ~Automate the process of sorting cucumbers in his parents farm. ~7000 pictures of the cucumbers, sorted by his mother, where used to train the dataset. ~Makoto used a softmax regression complemented by a convolutional network. He used TensorFlow to apply the model. ~After completion, Makoto validated his model with the test image and obtained an accuracy of 95%, however there is room for improvement. He suspects the accuracy is so high because the model is overfitting. Furthermore, the model can only classify on a few parameters, such as shape, length and level of distortion leaving other characteristics such as colour, texture and scratches out of the equation. ~The idea of the problem is not very different from identifying digits. I like how he used it to improve a very manual method. Would be interesting to use computer vision to analyse texture and scratches on the surface of the cucumbers and add it to the training model. 
74 ~63 ~https://www.sciencedaily.com/releases/2017/11/171115091819.htm ~University College London ~Wanted to determine whether machine learning could improve the ability to determine whether new drugs have affects on the brain ~Large scale data set from stroke patients, including anatomical patterns of brain damage ~Used gaze direction as indication of stroke impact and simulated meta-analysis of hypothetical drugs ~The researchers discovered that machine learning based models may allow better analysis of new drugs for the brain compared to statistical models ~
149 ~91 ~https://www.rdmag.com/article/2017/11/research-team-wins-award-machine-learning-diagnostic ~A team from Sandia National Laboratories and Boston University ~Develop a system that could perform diagnostic tests on supercomputers ~The team compiled a collection of issues they had come across in their time working with supercomputers and then  created code that would induce those  problems. Next they ran a variety of programs on computers with and without  their anomalous code. They collected information on the energy, processor and memory used and this was provided to their  algorithms to learn from. ~One of the techniques used was Random Forest which was highly effective at processing the vast number of data points.  This was later streamlined further by using averages, fifth percentile and 95th percentile in certain calculations. ~Resulted in a program that could use less than 1% of the systems processing power to analyse data and fix these  problems in the background. ~
73 ~110 ~https://www.newscientist.com/article/2146703-even-a-mask-wont-hide-you-from-the-latest-face-recognition-tech/ ~Amarjot Singh at the University of Cambridge and his colleagues ~a machine learning algorithm to identify a face when wearing a disguise. ~hand-labelled 2000 photos of people wearing hats, glasses, scarves and fake beards ~none ~The system accurately identified people a wearing scarf 77 per cent of the time, a cap and scarf 69 per cent of the time and a cap, scarf and glasses 55 per cent of the time. ~none 
32 ~112 ~https://techcrunch.com/2017/11/08/skype-launches-photo-effects-sticker-suggestions-powered-by-machine-learning/ ~Microsoft ~Detect faces in photos, determine subjects age and emotion, figure out celebrity look-a-like and suggest captions ~Not Mentioned ~Not mentioned ~Successfully released and users say app works as advertised ~N/A 
662 ~90 ~https://www.thewebmaster.com/seo/2016/apr/5/spam-algorithms-incorporate-machine-learning/ ~Agent was not specified in this study ~To provide a comprehensive machine learning algorithms comparison within the Web spam detection community. ~Two public web spam datasets were available, WEBSPAM-UK2006 and WEBSPAM-UK2007, both datasets provided 2 evaluated sets, set 1 was used for training and set 2 was used for testing ~Support Vector Machine - SVM 19 discriminates a set of high-dimension features using a or sets of hyperplanes that gives the largest minimum distance to separates all data points among classes. Multilayer Perceptron Neural Network - MLP 29 is a non-linear feed-forward network model which maps a set of inputs x onto a set of outputs y using multi weights connections. Bayesian Network - A BN 26 is a probabilistic graphical model for reasoning under uncertainty, where the nodes represent discrete or continuous variables and the links represent the relationships between them. Decision Tree - DT 39 decides the target class of a new sample based on selected features from available data using the concept of information entropy. The nodes of the tree are the attributes, each branch of the tree represents a possible decision and the end nodes or leaves are the classes. Random Forest - RF 13 works by constructing multiple decision trees on various sub-samples of the datasets and output the class that appear most often or mean predictions of the decision trees. Nave Bayes - The Nave Bayes 41 classifier is a classification algorithm based on Bayes theorem with strong independent assumptions between features. K-nearest Neighbour - KNN 2 is an instance-based learning algorithm that store all available data points and classifies the new data points based on similarity measure such as distance. Adaptive Boosting (AdaBoost) - The weights of incorrectly labelled data points are adjusted in AdaBoost such that the following classifiers focus more on incorrectly labelled or difficult cases 24. LogitBoost - LogitBoost 25 is actually an extension of AdaBoost where it applies the cost function logistic regression to AdaBoost, thus it classifies by using a regression scheme as base learner. Real AdaBoost - Unlike most Boosting algorithms which returns binary valued classes (Discrete AdaBoost), Real AdaBoost 42 outputs a real valued probability of the class. Bagging - Bagging 12 is a method by generating several training sets of the same size and use the same machine learning algorithm to build model of them and combine the predictions by averaging. It is often improve the accuracy and stability of the classifier. Dagging - Dagging 45 generates a number of disjoint and stratified folds out of the data and feeds each chunk of data to a copy of the machine learning classifier. Majority vote is done for predictions since all the generated machine learning classifier are put into the voted Meta classifier. Dagging is useful for base classifiers that are quadratic or worse in time behaviour on the number of instances in the training data. Rotation Forest - The rotation forest 40 is constructed using a number of the same machine learning classifier typically decision tree independently and trained on a new set of trained features form by sub-sampling of the datasets with principal component analysis applied on each sub-sets. ~Web Spam detection is also known as binary classification problem (spam or non-spam), thus the area under receiver operating characteristic curve (AUC) is used as evaluation metrics. The receiver characteristic curve is determined by plotting true positive rate vs the false positive rate in various threshold value. AUC is a measure for accuracy and also a performance metric for logistic regression. Most algorithms performed well with the idnedifcation of spam, with a rate of between 60% and 80%, however  random forest has outperform other classifiers including SVM which widely used in Web spam community as much as  0.927 in WEBSPAM-UK2006 and 0.850 in WEBSPAM-UK2007 AUC ~Since this study was performed, google has taken more steps to ensure the proper filteration of spam, and claims that Gmail can correctly detect 99% of all spam emails 
124 ~130 ~https://thenextweb.com/artificial-intelligence/2017/11/13/alibabas-ai-is-the-blueprint-for-brick-and-mortar-stores-of-the-future/ ~Suggest similar fashion options to the user based on what they have already looked at/bought. System is designed to access huge quantities of user data and make smart decisions as the user browses more. ~Alibaba ~User purchases, preferred fashion themes, site browse history, popular accessories to chosen clothes, user sizes. ~As user browses, the learning algorithm improves the suggestions of similar fashion, what this can be compared to and which items will end up being bought. ~No hard evidence. COMMENT: Alibaba implemented this system temporarily as an experiment for their 'Saturday Singles Day' event. This event generated $25 bn in sales, quadruple US spending for the last black Friday. We can assume that the implemented system helped somewhat with these staggering sales figures. 
56 ~123 ~https://techcrunch.com/2017/11/01/audibles-new-romance-audiobooks-service-uses-machine-learning-to-jump-to-the-sex-scenes/ ~Audible  ~To find the "good parts" of the audiobooks. ~Audible Romance audiobooks ~Audible is using ML to identify keywords and phrases, in order to figure out where things get hot and heavy.  ~Audible ML can identify 32 of romance sub-genres and 122 story and character imagery. ~It can ultimately turn into audiobook summaries. 
70 ~98 ~https://medium.com/artists-and-machine-intelligence/neural-nets-for-generating-music-f46dffac21c0 ~Ianniz Xenakis, for his 1958 compositions,  describing his process in  "Formalized Music: Thought and Mathematics in compositions ~To algorithmically generate musical compositions ~A database of Musical compositions and instrumental data ~Markov Chains, used to model probabilstic systems, take advantage of existing material. ~Iannz Xenakis managed to compose his own music algorithmically, this was before the time of computational machines, now this can be extended even further. ~None 
155 ~63 ~https://www.sciencedaily.com/releases/2017/11/171115091819.htm ~A team out of University College London lead by Dr Parashkev Nachev. ~To improve our ability to determine whether a new drug works in the brain of a human. ~The team looked at large scale data from patients who had suffered strokes and created the largest collection of  anatomically registered images of strokes ever assembled. ~They simulated a large-scale meta-analysis of a set of hypothetical drugs, to see if treatment effects of different  magnitudes that would have been missed by conventional statistical analysis could be identified with machine learning.  No algorithm given. ~Results of note were in relation to drugs that are used to shrink lesions on the brain. Conventional low-dimension  models required the lesion to shrink by 78.4% of its volume for the effect to be detected in trials more often than not.  With the high-dimensional modelling of their approach this threshold was reduced to 55%.  ~
191 ~265 ~Engadget https://l.facebook.com/l.php?u=https%3A%2F%2Fwww.engadget.com%2F2015%2F06%2F17%2Fsuper-mario-world-self-learning-ai%2F&h=ATPSzmDQwiuinWe4J3VFBed94zJ7GjJqCWPbJJoqHwXdyxc7a3XNwIM4OlRHGWcF581zhJhafYij34GkhKJ_dBbs2cbelC3z_B4-wxW6VVbaLrn_WXQjYKiqa85tIOM29MuxlQ ~Seth Bling ~Have the computer program learn the Mario level. ~The input that MarI/O, name of program, sees. A simplified view of the level, the white squares are blocks the program can stand on and the black squares are moving objects, such as enemies. ~The AI has a "fitness" level, which increases the further right the character reaches, and decreases when moving left. The AI knows that fitness is good, and so, once it figures out that moving right increases that stat, it's incentivized to continue doing so. Every generation introduced new ideas, but it was simply trying different things, not doing what it "thought" would work. When an idea was a success, it was remembered, when it wasn't, it was discarded and learned from. This learning style is called NeuroEvolution of Augmenting Topologies. ~The Program was able to complete the program in 34 attempts. ~This is a very quick evolution of the program to be able to walk in knowing nothing to being able to do it, without fault. May be able to be used in making better AI deal with situations in computer games in the future. 
62 ~94 ~https://www.outerplaces.com/science/item/16912-intel-machine-learning-artificial-intelligence ~Intel ~To integrate AI, machine learning and deep learning into a dedicated chip ~N/A METHOD Nervada Neural Network Processor (NNP). The chip has storage for deep learning. ~N/A ~Having a chip specifically for AI and machine learning would be of great benefit for the field. However, I share the skepticism of the article with regards to Intel's purchase of Nervana Systems.   
412 ~104 ~https://www.theverge.com/2016/6/1/11829678/google-magenta-melody-art-generative-artificial-intelligence ~The agent here is Google. ~The goal is to use machine learning techniques in order to teach a system on how to compose a piece of music, and on a deeper level to advance the state of the art in machine intelligence for music and art generation. Google also hope to build a community of artists, coders and machine learning researchers. The system collected is called Magenta. ~Magenta uses the NSynth Dataset. NSynth is an audio dataset containing 305,979 musical notes, each with a unique pitch, timbre, and envelope.For 1,006 instruments from commercial sample libraries, Google generated four second, monophonic 16kHz audio snippets, referred to as notes, by ranging over every pitch of a standard MIDI pian o (21-108) as well as five different velocities (25, 50, 75, 100, 127).  ~Magenta is build upon Google's own TensorFlow system. The NSynth (Neural Audio Synthesis) algorithm was used in this project, googles own new approach to synthesis using neural networks. NSynth works by finding a compressed representation of sound(we will call Gamma). An encoder network transformers a sound into is Gamma representation. A decoder network then converts it back to sound. The system is trained in such a way that the reproduced sound is as similar as possible to the real sound.  ~While regularly we could measure likelihood for the success of similar machine learning projects, this cannot be applied as comprehensively to music and art due to their inherently subjective nature. On the note of this subjective nature, I believe Magenta has produced fantastic results in terms of some of the music it has produced. The amount of articles concerning and even YouTube recordings of Magentas efforts further this point.  ~ I think this is an extrodinary project when mapped over time. Some examples of the music produced by Magenta are quite honestly poor, however there are more advanced demos that are extraordinary. It was once said that as advanced computers could be, they could not be artistic or creative. Projects like magenta shatter this idea - and even still are in their relative infancy. It would be interesting to see what this amount to in, say, 20 years time. Will we be attending concerts and festivals headed by famous composer-algorithms? Will these endeavours eventually produce music so high quality that human beings cannot create something better? With more research and advancements, this sort of application of Machine Learning has the potential to be game-changing for the music scene.  
175 ~132 ~S. Roberts, M. Osbourne, M. Ebden, S. Reece, N. Gibson, S. Aigrain http://rsta.royalsocietypublishing.org/content/371/1984/20110550 ~The Royal Society publishing. publishes mathematical, physical and engineering sciences papers and articles. ~To introduce the Gaussian processes for time-series data analysis and the conceptual framework of Bayesian modelling in this filed.  ~ dataset D = {(x1, y1),(x2, y2)} consist- functions that agree ing of two observations ~GPs - Gaussian processes provide a principal practical aproach to learning in karnel machines.Training the GP model involves both model selection, or the discrete choice between different functional forms for mean and covariance functions as well as adaptation of the hyperparameters of these function. n the Bayesian approach, all unknown quantities are considered as random variables and uncertainties over those quantities are represented using probability distributions conditional on the available data  ~Gaussian processes advantages with respect to the interpretation of model predictions and provides a wellfoundedframework for learning and model selection. Theoretical and practical developments of over the last decade have made Gaussian processes a serious competitor for real supervised learning applications. ~
33 ~36 ~https://www.deepinstinct.com/#/home ~Deep Instinct ~Detect potential malware ~Continuously updated set of both malicious and benign files ~Data Traning and pattern matching ~Accuracy for files with 2 to 10 percent difference to older versions ~
147 ~91 ~Article on ScienceDaily.com https://www.sciencedaily.com/releases/2017/11/171115091819.htm ~University College London ~Use machine learning to improve our ability to determine whether a new drug works in the brain. ~The research team looked at large-scale data from patients with stroke, extracting the complex anatomical pattern of brain damage caused by the stroke in each patient. ~The machine learning algorithm took into account the presence or absence of damage across the entire brain, treating the stroke as a complex "fingerprint", described by a multitude of variables. ~Conventional methods would only detect an effect in a trial if the lesion was shrunk by 78.4% of its volume, while the machine learning method would detect an effect when the lesion was shrunk by only 55%. ~The researchers say their finding demonstrate that machine learning could be invaluable to medical science, especially when the system under study - such as the brain - is highly complex. 
41 ~130 ~https://thenextweb.com/artificial-intelligence/2017/11/13/alibabas-ai-is-the-blueprint-for-brick-and-mortar-stores-of-the-future/ ~Alibaba's fashionAI ~To make predictive guesses to help people using the AI get recommendations for what clothes the user is interested in and help make fashion tips for the user. ~accessing massive quantities of data in real-time. ~NA ~NA ~NA 
63 ~94 ~https://news.stanford.edu/2017/11/15/algorithm-outperforms-radiologists-diagnosing-pneumonia/ ~Stanford researchers  ~An algorithm that offers diagnoses based off chest X-ray images ~A dataset contains 112,120 frontal-view chest X-ray images labeled with up to 14 possible pathologies ~420 of the images were annotated for the test set. They then used a training algorithm. ~They created an algorithm which exceeds the perfor-mance of radiologists in detecting pneumonia fromfrontal-view chest X-ray images. ~N/A 
64 ~69 ~http://www.iflscience.com/technology/artificial-intelligence-dreams/ ~Google Engineers ~To test how quickly and effectively computers are learning. ~Huge database of random classified images. ~The optimization resembles Backpropagation, however instead of adjusting the network weights, the weights are held fixed and the input is adjusted. ~Generation of psychedelic and surreal images. ~The software is designed to detect faces and other patterns in images, with the aim of automatically classifying images. 
118 ~114 ~Online article on Engadget https://www.engadget.com/2017/08/01/australian-budget-bot-wins-amazon-robot-challenge/ ~A group of Australian robotics students ~To construct a robot that can accurately pick and stow items from bins to boxes in a short time. ~Various images of items were gathered by the students. These images were preprocessed using algorithms and methods. ~Computer vision, artificial intelligence, neural networks. ~The Australian team won the Amazon Robotics Challenge with a robot that stood above the rest. ~I have had an opportunity to join a team for this contest at Texas State. This is a rigorous event which Amazon holds every year. It really helps to have a deep background of machine learning and engineering to construct an effective robot to identify, pick, and stow items. 
139 ~70 ~http://news.mit.edu/2016/faster-automatic-bug-repair-code-errors-0129 ~ Fan Long,  Martin Rinard ~ Creation of a automatic bug removal tool ~ errornous code (bugs) faced by users of open source project and the solution to the bugs posted by them ~ recognizing various values stored in memory during runtime like variables and constants. They identified 30 prime characteristics of a given value and that evaluated all the possible relationships between these characteristics in successive lines of code. More than 3,500 such relationships constitute their feature set. Their machine-learning algorithm then tried to determine what combination of features most consistently predicted the success of a patch ~ they were able to fix 10 times as many errors as its predecessors ~ this could help in the final stage of the lifecycle of a program where most of the time is spend during testing and debugging. 
60 ~84 ~https://sciencebusiness.net/network-news/improving-clinical-trials-machine-learning ~Wellcome and the National Institute for Health Research University College London Hospitals Biomedical Research Centre. ~Improve clinical trials with machine learning ~Image Collection of the anatomical pattern of brain damage from strokes. ~Generic Machine Learning, didn't actually say. ~Machine learning method particularly strong compared to statistical models when looking at interventions that reduce the volume of a lesion. ~
200 ~125 ~Article on PHYS.ORG by Aaron Dubrow https://phys.org/news/2017-10-scientists-supercomputers-machine-automatically-brain.html ~George Biros, professor of mechanical engineering at The University of Texas at Austin, along with collaborators from the University of Pennsylvania, University of Houston, and University of Stuttgart. ~The goal was to create a system which could characterize gliomas, the most common and aggressive type of primary brain tumor. ~The system was calibrated with training data consisting of 300 sets of brain images. The system was then then tested on data from 140 patients. ~The system has two main steps: a supervised machine learning step where the computer creates a probability map for the target classes, and a second step where they combine these probabilities with a biophysical model that represents how tumors grow in mathematical terms, which imposes limits on the analyses and helps find correlations. ~Biros and his team were able to run their analysis pipeline on 140 brains in less than 4 hours and correctly characterized the testing data with nearly 90 percent accuracy, which is comparable to human radiologists. ~The methods that the team developed go beyond brain tumour identification. They are applicable to many problems in medicine as well as in physics, including semiconductor design and plasma dynamics. 
124 ~59 ~https://newatlas.com/open-ai-dota2-machine-learning/50882/ ~OpenAI ~To develop a bot that can defeat top players in the popular game Dota 2. ~The data was optained by continually playing the bot against a copy of itself. ~OpenAI set the bot to teach itself how to play the game through self-play. The system learned to conquer the game from scratch by playing a mirror of itself. ~After just two weeks of learning, the bot beat several of the world's top Dota 2 players. At this stage the bot only plays in the more simplistic one-to-one version of Dota 2. The full, and exponentially more complex, version is played by two teams of five. The OpenAI team are now working on teaching teams of bots to play this complete version. ~
256 ~99 ~https://towardsdatascience.com/background-removal-with-deep-learning-c4f2104b3157 By Gidi Shperber ~Self developed project. Unreleased. ~Background removal with deep learning. When given a 2D picture representation of anything the application should be able to find the main object in the picture and cut out any background surrounding it. ~The most common datasets for segmentation were the COCO dataset, which includes  around 80K images with 90 categories, the VOC pascal dataset with 11K images and  20 classes, and the newer ADE20K datasets. The COCO dataset was chosen as the main source of data since it includes  much more images with the class “person” which was our class of interest. ~The model was trained with a schedule: standard cross  entropy loss, RMSProp optimizer with 1e-3 learning rate and small decay. The 11,000 images were split into 70% training, 20% validation, 10% test.  To keep to the training schedule aligned with the original paper, the epoch size was set on 500 images. This also allowed the model to be saved periodically with every improvement in results,  Additionally, the model was trained on only 2 classes: background and person. ~The results show a IoU value of 84.6 on the test set, while current state of the art is 85.  That number fluctuates throughout different datasets and classes.  There are classes which are inherently easier to segment e.g houses, roads, where most models  easily reach results of 90 IoU. Other more challenging classes are trees and humans,  on which most models reach results of around 60 IoU. ~
64 ~143 ~ https://www.forbes.com/sites/omribarzilay/2017/11/14/trueaccord-nabs-22m-series-b-to-bring-machine-learning-to-debt-collections/#112c6d1b7f09 ~TrueAccord ~The goal is to disrupt the debt collection industry. ~They used their 1.8 million customers as data. METHOD They used a machine learning algorithm to analyse customer trends. They then tailored their debt collection methods to each customer based off what was successful with similar customers. RESULT The algorithm had a minimum increase of 50% (to 500%) in debt collection rates. 
30 ~94 ~https://news.stanford.edu/2017/11/15/algorithm-outperforms-radiologists-diagnosing-pneumonia/ ~Taylor Kubota ~Discuss deep learning algorithm that identifies pneumonia from chest xrays ~Dataset of front view chest xrays ~None ~Outperformed radiologists at identifying pneumonia after about one month ~None 
120 ~114 ~https://www.newscientist.com/article/mg23231020-100-ai-learns-to-predict-the-future-by-watching-2-million-videos/ ~Carl Vondrick ~They are showing a little passage from a video and trying to make the machine predict what happens next. ~The team trained the algorithm with 2 million videos from Flickr.These videos were unlabelled. METHOD Adversarial Networks were used for this. One of them was generating the videos and the other one was trying to judge whether if it is fake or not. ~The results are still blurry. Since they are still developing it.For now,according to Vondrick his program is still like a 10 years old child. ~I have to admit, it sounds pretty exiciting to think the program as a child. Every development and every experience it gets, it is actually kind of growing like a child. 
116 ~15 ~BioInformatics ~Pierre baldi and Soren Bruna ~Using the essential tools for rapid growth of bio informatics field. Building computer systems that can adapt to environments and learn from experience has attracted researchers from many field. ~Biological data required and automated analysis. Generous of many important species have been completely sequenced, new technologies of DNA and protein array have been practical research tools. ~Bayesian reasoning,  Maximum entropy methods, Gaussian processes and support vector machines. ~Full genome sequencing has blossomed with the completion of the sequence of the fly and the draft of the human genome project. ~The improvement of bioinformatics is essential in the important face of technology and machine learning will have an greater impact. 
130 ~98 ~http://news.mit.edu/2017/artificial-intelligence-for-your-blind-spot-mit-csail-cornercameras-1009 ~Adam Conner-Simons  ~An algorithm for rectifying blind spot, using smartphone cameras, system for seeing around corners could help with self-driving cars and in search-and-rescue missions. ~Data gathered from cameras, sensing systems etc. like lights by the edge of the wall, reflections of the objects around the camera.  ~Analyzing light at the edge of walls, which is impacted by the reflection of objects around the corner from the camera. ~imaging system, which can work with smartphone cameras, uses information about light reflections to detect objects or people in a hidden scene and measure their speed and trajectory all in real-time. ~A really use full system with many real life applications like for eg: giving drivers warning a few seconds ahead in time before which can save many lives. 
62 ~55 ~CORDIS http://cordis.europa.eu/news/rcn/128663_en.html ~Prof Alessandro Bogliolo from the University of Urbino. ~Detect road surface anomalies. ~Smart phone app gathers accelerometer and gyroscope data. ~Supervised Machine Learning. ~An open data set with more than 2 000 000 records, automatically updated every 6 hours with crowd-sensed data, has been published. ~Could optimise how we choose roads to repair improving the efficiency of the Council services. 
137 ~128 ~Business Insider http://uk.businessinsider.com/visualdx-machine-learning-app-for-skin-diagnosis-ceo-interview-2017-11?r=US&IR=T ~Tim Cook & Apple ~Enable doctors to be able to take a photo and identify the disease or condition depicted ~Many of VisualDx's images were scanned from old collections of slides and film, from leading departments and doctors. It's built a library of 32,000 images to train its models. "We developed this reputation as somebody that was going to preserve the legacy of medical photography," ~Doctors can take a photo on an iPad or iPhone and then it will locally analyse the image against the images in the database to find the closest match ~The results of this development is an app that can use a user inputted image to help diagnosing skin conditions. The result is an accurate way of detecting medical issues. ~This app has been developed to help diagnose skin conditions. 
109 ~114 ~https://siliconangle.com/blog/2017/11/12/machine-learning-startup-graphcore-raises-50m-round-led-sequoia-capital/ ~Graphcore LTD. ~Building intelligence processor chips that are specifically designed to assist programmers in creating machine learning systems. ~Due to the wide range of applications for a system like this, there is a gargantuan amount of data that could be collected spanning over all applications. ~The company claims its IPU accelerators and poplar software framework deliver the fastest and most flexible platform for current and future machine intelligence applications, lowering the cost of AI in the cloud and datacenter. Also improving performance and efficiency by between 10x and 100x. ~None. ~Graphcore have recently secured a $50 million investment which shows the incredible potential these intelligence processor chips have. 
180 ~109 ~http://blog.kaggle.com/2017/03/24/leaf-classification-competition-1st-place-winners-interview-ivan-sosnovik/ ~Ivan Sosnovik ~Classify leaves according to their features and use the model to compete in Kaggle. ~Consists of 1584 images of leaf specimens (16 samples each of 99 species), converted into binary black leaves agains white backgrounds. Three sets of features were also provided, shape contiguous descriptio, interior texture histogram and a fine-scale margin histogram. ~Ivan first implemented a PCA with 35 components, which gave good results for most of the objects. However, a few objects had a lot of uncertainty on their class. Ivan decided to look at these objects and found that the distinction between them was very fine and even he would have trouble classifying them. To solve this he decided to implement a Random Forest classifier for those objects with high uncertainty. Thus the final classifcation included a mix of objects classified by PCA and Random Forest. ~With the two classifiers, Ivan was able to correctly identify all leaves in the dataset and win the competition. ~Kaggle is a great learning environment for data science. Data is nicely provided and the projects are well defined. 
67 ~92 ~Google Research: Email Category Prediction (https://research.google.com/pubs/pub45896.html) ~Aston Zhang, Luis Garcia Pueyo, James B. Wendt, Marc Najork, Andrei Broder ~Compared the performance of Neural Network and Markov Chain to categorize emails. ~Email category sequences spanning 90 days of activity for 100,000 anonymized email users. ~Markov Chain, Long short-term memory and Multilayer Perceptron. ~MLP is far ore effective than Markov Chain and LSTM is slightly better than MLP. ~
102 ~96 ~https://futurism.com/ibms-watson-ai-recommends-same-treatment-as-doctors-in-99-of-cancer-cases/ ~IBM’s Watson & Human experts at the University of North Carolina School of Medicine  ~To contribute to the healthcare industry and medical researches specially oncology.  ~Nearly 1,000 cancer diagnoses cases. ~AI to analyze 1,000 cancer diagnoses and recommend treatments accordingly. ~Watson was able to recommend the same treatment which are the actual treatments from oncologists in 99% of the cases. Also because it can read and analyze thousands of medical documents in minutes, Watson found treatment options human doctors misplaced in 30% of the cases. ~It is promising that artificial intelligence is put in good use for medical purposes. 
27 ~71 ~http://news.mit.edu/2015/learning-language-playing-computer-games-0924 ~MIT Researchers ~Learning a language by learning to play a text based game. ~none ~Deep Learning ~Outperformed "bag of words" and "bag of bigrams" techniques. ~none 
95 ~128 ~https://www.sas.com/en_us/insights/articles/analytics/data-for-good-protecting-consumers-from-unfair-financial-practices.html#/ ~Tom Sabo ~He wanted to find a way assess the data for various trends and discover the areas of biggest concern for custumers. ~37.000 complaints were used for his analysis. ~He used text analytics to explore sentiment in complaints and used machine learning for modelling the natural language available ~As a result of this process,complaints started to decline. ~Fatigue is a huge problem for those who has to read all those complaints and after all those documents they can make mistakes and missed the required spots as a feedbacks.Considering this, it is a neccessity. 
152 ~64 ~http://jmlr.org/papers/volume15/srivastava14a/srivastava14a.pdf ~Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, Ruslan Salakhutdinov ~Large networks are also slow to use, making it difficult to deal with overfitting by combining the predictions of many different large neural nets at test time. Dropout is a technique for addressing this problem. ~Randomly drop units (along with their connections) from the neural network during training. ~They randomly dropped units from the neural network during training, preventing co-adaption too much. ~Random dropout breaks up these co-adaptations by making the presence of any particular hidden unit unreliable. This technique was found to improve the performance of neural nets in a wide variety of application domains including object classification, digit recognition, speech recognition, document classification and analysis of computational biology data. This suggests that dropout is a general technique and is not specific to any domain. Methods that use dropout achieve state-of-the-art results on SVHN, ImageNet, CIFAR-100 and MNIST.  ~
149 ~110 ~- Paper titled - The Learning Behind Gmail Priority Inbox by Douglas Aberdeen ,Ondrej Pacovsky, Andrew Slater ~- Douglas Aberdeen ,Ondrej Pacovsky, Andrew Slater ~-  The Priority Inbox attempts to alleviate such information overload by learning a per-user statistical model of importance, and ranking mail by how likely the user is to act on that mail ~- how the user interacts with a mail after delivery ~- Linear logistic regression -  the final prediction is the sum of the global model and the user model log odds. Mails are classified in to two categories they are important and not important. Opening a mail is a strong signal of importance here. ~- The success rate was around 80-85% ~- Since this technique is based on wheather the user open a mail or not, many user tend to open not so important mails, which led decrease in success rate 
89 ~151 ~New York University https://www.nyu.edu/about/news-publications/news/2017/august/researchers-use-machine-learning-to-spot-counterfeit-consumer-pr.html ~Team of researchers led by New York University Professor Lakshminarayanan Subramanian. ~A non-intrusive solution to easily distinguish authentic versions of the product produced by the original manufacturer and fake versions of the product produced by counterfeiters. ~A dataset of three million images across various objects and materials such as fabrics, leather, pills, electronics, toys and shoes.  ~Not mentioned. ~The solution was able to classify counterfeit and authentic products with 98% accuracy. ~If utilised during customs procedures it could make importing counterfeit goods more difficult. 
96 ~102 ~Machine Learning Methods for Strategy Research (http://www.hbs.edu/faculty/Pages/item.aspx?num=53076) ~Mike Horia Teodorescu ~Natural language processing methods focused on text analytics and machine learning methods with their applications to management research and strategic practice. ~A set of standard machine learning techniques with a view towards management research and strategic practice. ~An accessible overview of machine learning techniques with examples and information on how to apply these to management. ~This article may help researchers and managers alike to master some of the tools in machine learning, particularly in using natural language processing methods, decision trees, clustering, and classification methods. ~
252 ~92 ~http://news.mit.edu/2017/artificial-intelligence-suggests-recipes-based-on-food-photos-0720 ~Pic2Recipe ~To identify the ingredients of a recipe given a picture of the prepared food. This project aimed to build off of this work but dramatically expand in scope, websites like All Recipes and Food.com to develop Recipe1M, a database of over 1 million recipes that were annotated with information about the ingredients in a wide range of dishes ~Swiss researchers created a food-101 dataset in 2014. it consists of a wide variety of pictures as well a recipes that link to those pictures. ~They data in the Recipe1m dataset was used to train a neural network to find patterns and make connections between the food images and the corresponding ingredients and recipes. ~The team’s approach works at a similar level to human judgement according to the paper, although it does not give actual percentages, the swiss team in 2014 developed an algorithm that could recognize images of food with 50 percent accuracy. Future iterations only improved accuracy to about 80 percent, and it was ruled that the dataset was the limiting factor, since Recipe1M is a much larger dataset, this may be the push needed to get past 80%. ~This study was effected by regional cuisine, for example the City University in Hong Kong has over 110,000 images and 65,000 recipes but it is all tied to Chinese cuisine, until a conslolidated dataset could be established, the use of this is very limited. It also does not acccount for home or family recipes, that can contain modifications to recipes easily. 
60 ~56 ~https://www.theengineer.co.uk/algorithm-robots-fastron/ ~Prof Michael Yip( Team Lead ), University of California San Diego, ~Develop faster collision detection algorithms for robots that operate in human environments. ~Sensor data from the robot. ~Fastron algorithm used to classify collisions versus non-collisions points. ~Fastron algorithm can run up to 8 times faster than traditional collision detection algorithms. ~Interesting use of machine learning in robot-assisted surgeries. 
137 ~143 ~https://www.newscientist.com/article/2130205-inquisitive-bot-asks-questions-to-test-your-understanding/ (https://arxiv.org/pdf/1705.00106.pdf) ~Xinya Du, Junru Shao, Claire Cardie ~Create a system that can generate questions to test the users comprehension of a giving piece of text.  ~NLP datasets such as 'SQuAD' and 'MS MARCO' ~Text is analysed on a sentence and sentence plus paragraph level. Given sentence-question pairs from the dataset the model tries to minimise the negative log likelihood of the training data with respect to all parameters ~This question generation model outperforms previous models that used human entered rules for generation, but performs worse than questions directly created by humans. ~Could be a useful tool for study as it will ensure comprehension of the material, and even though it performed worse than a human it is still very useful as it can generate questions on a much larger scale than a human possibly could. 
177 ~107 ~https://www.geekwire.com/2016/uber-collapse-without-pattern-finding-computers-says-chief-machine-learning/ ~Uber ~Wanted to improve their app's ability to estimate times for pickups and deliveries. ~Measured their estimated times and actual times. They then compared the two to see how accurate their estimates really were. ~Initially, (in the case of Uber Eats), they computed the time using the distance between the customer and the restaurant, the average speed, and the time required to prepare the meal. They upgraded this by having their app assess their collected data of real world (not estimated) delivery times for several thousand previously deliveries, and making a prediction based off of that. ~Uber continued to compare their estimates to the actual devilery times. According to Danny Lange; head of machine learning at Uber, their estimate times improved by 26 per cent after the addition of machine learning to their prediction algorithm. ~A interesting case, certainly not one I was expecting to find while researching this topic. However, it does make a lot of sense, and I would imagine a lot of similar companies are also making a similar use of machine learning. 
249 ~192 ~ https://www.theregister.co.uk/2017/07/18/knightscope_k5_falls_into_pond/ https://www.knightscope.com/demo https://en.wikipedia.org/wiki/Knightscope https://www.knightscope.com/knightscope-k5 ~Knightscope K5 was developed by Knightscope Inc. Knightscope Inc is a security company that builds Autonomous Data Machines to monitor shopping, parking lots and neighbourhoods.  ~ K5's goal is to survey its surroundings and to "detect crime". The model was developed in 2013, and is the oldest robot developed by Knightscope Inc. It's duties include contacting the authorities should it detect abnormal changes in data received from its sensors and cameras. ~ The K5 creates a 3-D image of its surroundings. It uses an ultrasonic sensor to detect objects in surroundings and monitors its own movements within its environment. ~ The machine uses a combination of cameras, thermal imaging sensors, radar, air quality sensors and a microphone to monitor its surroundings. The machine can learn how to react to situations depending on previously obtained data from its surroundings. Over time the robot can determine what is "normal" and will create an alert if it deems any data "irregular". ~ During the patrolling duties of one K5 robot tragedy struck. The machine was unable to detect the steps leading to a water feature in The Washington Harbour retail complex.  The same model had previously been the public eye for knocking over a 16 month old child. ~ The K5 appears to either suffer from the dreaded outdated curse, or there exists a lack of interest with regards to developing the model. There may be more appearances of Knightscope Inc. in the tabloids yet. 
89 ~267 ~http://www.businessinsider.com/visualdx-machine-learning-app-for-skin-diagnosis-ceo-interview-2017-11/?r=US&IR=T&IR=T/#visualdx-is-intended-for-use-by-doctors-to-confirm-and-validate-diagnoses-it-allows-doctors-to-search-by-symptoms-signs-and-other-patient-factors-1 ~New York based company called VisualDx ~To use machine learning model in a mobile application to help in the diagnosis of different skin conditions by taking pictures of skin. ~They collected 32,000 images by scanning old collections of slides and film, from departments and doctors. ~Deep neural networks were trained to recognise different skin conditions. ~Apple has recognised the usefulness for doctors, but it's not good enough as a diagnosis tool for the public.  ~VisualDx CEO Art Papier doesn't think completely automated diagnosis will happen anytime soon. 
106 ~96 ~https://www.wired.com/story/ai-can-help-apple-watch-predict-high-blood-pressure-sleep-apnea/amp ~Cardiogram and University of California, San Francisco. ~To find out if data from Apple Watch can be used to predict hypertension or sleep apnea.  ~Heart-rate and step count data from Apple Watches of over 6,000 volunteers combined with information about their health. ~They trained the kinds of artificial neural networks that are often used in speech recognition systems. ~With only one week of data from the wearer, their algorithm can predict hypertension with an accuracy of about 80 percent, and sleep apnea with an accuracy of about 90. ~More testing needs to be done to fully evaluate the usefulness of the findings in medicine. 
43 ~63 ~https://www.sciencedaily.com/releases/2017/11/171115091819.htm ~University College London ~To find out if machine learning could improve the ability to determine whether a new drug works in the brain ~Data from patients with a stroke ~Simulated a large scale meta analysis of a set of hypothetical drugs ~None 
140 ~133 ~Technology Review https://www.technologyreview.com/the-download/609510/a-new-algorithm-can-spot-pneumonia-better-than-a-radiologist/ ~A research from Stanford Pranav Rajpurkar, Jeremy Irvin, Kaylie Zhu, Brandon Yang, Hershel Mehta, Tony Duan, Daisy Ding, Aarti Bagul, Curtis Langlotz, Katie Shpanskaya, Matthew P. Lungren, Andrew Y. Ng ~Using machine learning to analyse x-rays and identify people with pneumonia correctly. ~ChestX-ray14 publicly available chest X-ray dataset containing over 100,000 frontal-view  X-ray images  with  14  diseases. ~CheXNet convolutional neural network  ~The researchers had four radiologists go through a test set of x-rays and make diagnoses, which were compared with diagnoses performed by CheXNet. Not only did CheXNet beat radiologists at spotting pneumonia, but once the algorithm was expanded, it proved better at identifying the other 13 diseases as well. ~Being able to outperform radiologists on diagnosing over 14 different diseases shows the tech is promising for diagnosing medical issues more accurately. 
128 ~118 ~Harvard Business Review https://hbr.org/2017/05/how-machine-learning-is-helping-us-predict-heart-disease-and-diabetes ~Yannis Paschalidis, Boston University’s Center for Information and Systems Engineering ~To Predict Heart Disease and Diabetes ~Patients’ anonymized electronic health records (EHRs) that contain all of the information the hospital has about each patient, including demographics, diagnoses, admissions, procedures, vital signs taken at doctor visits, medications prescribed, and lab results. ~Their algorithms predict who might have to be hospitalized. This gives the hospital opportunities to intervene, treat the disease more aggressively in an outpatient setting, and avoid a costly hospitalization while improving the patient’s condition ~They could predict hospitalizations due to these two chronic diseases about a year in advance with an accuracy rate of as much as 82%. ~This could both improve healthcare and the ecnomy saving billions of dollars/euros by preventing hospitalizations. 
138 ~96 ~https://betakit.com/how-quantum-machine-learning-will-solve-problems-once-thought-out-of-reach/ ~Creative Destsruction Lab ~To solve difficult problems with the use of "Quantum machine learning". ~N/A ~Quantum computers are used which uses machine learning to solve problems. The data is prepared with a method called "quantum state preperation" followed by deciding on a language to use. The last step involves a step called "measurement" is used to be able to read the results ~Various results including one from Switzerland, specifically. It took a few days for a computer with "100 to 200 quantum bits" to output the result for difficult chemical reaction. ~I wonder wether the kinds of problems could be extended to the likes of P vs. NP, or even hard mathematical problems? Applying AI and machine learning algorithms on a quantum computer is fascinating and could perhaps deal with questions in philosophy such as determinism etc. 
106 ~102 ~Online article on Engadget https://www.engadget.com/2017/06/27/microsoft-sports-performance-platform/ ~Microsoft ~To help teams track and improve player behavior in sports. ~Athlete statistics, team statistics on wins and losses, etc. ~Big data manipulation with machine learning. ~It is an experimental project for Microsoft. The company looks for sports organizations to be a part of the project. The ultimate goal is to use machine learning to help improve teams in sports. ~This is a good example of using machine learning in sports. It would be great to use classifying algorithms to determine if a player is at risk of injury during play. This may save money and time for players and organizations. 
130 ~122 ~SOURCE:https://www.technologyreview.com/s/602958/an-ai-ophthalmologist-shows-how-machine-learning-may-transform-medicine/ ~Google and Ophthalmologists (NHS) ~To detect eye diseases such as diabetic retinopathy ~They used patients to observe their retinae and captured a picture with a specialized device for signs of bleeding and fluid leakage.  ~Automated detection made the diagnosis more efficient and reliable. They analysed retinal images. It looked at thousands of healthy and diseased eyes and figured out for itself how to spot the condition ~They managed to find an algorithm which detects the above-mentioned disease as well as a highly trained ophthalmologist can. ~If this keeps improving, we will have more accurate diagnoses which will help detect many more eye diseases. These will be more efficient than a human doctor. Assuming the results are accurate. This could help people where certain doctors are not readily available. 
189 ~134 ~ http://www.digitaljournal.com/tech-and-science/technology/john-deere-advancing-machine-learning-in-agriculture-sector/article/502194 ~ John Deere ~ See & Spray Weed Killing technology - To detect and identify individual plants and make decisions on how to treat each plant individually. Reducing the use of herbicides, & allowing the replacement of some herbicides with organic ones, thereby reducing monetary cost, reducing impact on environment, reducing the levels of herbicides making it into the food chain and reducing the likelihood of weeds growing resistance to herbicides. Crop yields would also be expected to improve. Historically weeds are treated on a grand scale, by the field, including the spraying of the desired crop with a general purpose weed killer which is designed to hit as many of the varieties of weeds as possible. This Machine Learning method will reduce herbicide usage by up to 95%. ~ None mentioned in the article other than first roll out was targeted at Cotton farmers first with Soybean farmers to be targeted next.  ~ None shown in the article. ~ Up to 95% savings on herbicide use quoted in articles reviewed.  ~System is planned for commercial launch in 2018. Further links of interest: https://www.deere.com/en/our-company/news-and-announcements/news-releases/2017/corporate/2017sep06-blue-river-technology/ http://smartmachines.bluerivertechnology.com/ 
359 ~279 ~ https://www.faceapp.com/ https://play.google.com/store/apps/details?id=io.faceapp&hl=en https://techcrunch.com/2017/04/25/faceapp-apologises-for-building-a-racist-ai/ http://www.telegraph.co.uk/technology/2017/04/25/faceapp-viral-selfie-app-racism-storm-hot-mode-lightens-skin/ ~ FaceApp is a mobile application that was developed by Wireless Lab.  ~ It uses neural networks to transform photographs of faces into very realistic alternate versions. The official website boasts that it can "transform your face using artificial intelligence with just one tap". The application allows the user to alter images to add a smile, change gender, look older, look younger and look "more attractive". The latter being the most controversial of all the filters. ~ FaceApp uses a combination of its own training data and open-source libraries including TensorFlow. Goncharov has been recorded as admitting that the training data for determining desirability was the company's own training data.  He has admitted that the training data had biases. "We are deeply sorry for this unquestionably serious issue. It is an unfortunate side-effect of the underlying neural network caused by the training set bias, not intended behaviour." ~ The applications used a combination of machine learning algorithms alongside photo editing algorithms. ~ The application received feedback from users stating that it was "racist" when it came to determining how "hotness" should be characterised. When used by black people, the filter automatically brightened the users' skin tone. The companies first relabelled the filter to "spark" instead of its previous name "hotness", and then later removed the filter completely. Wireless Lab claim to be currently working on a fix for this filter. COMMENT I don't feel that Wire Lab immediately learned and grew from their negative publicity due to other controversies regarding claims of racism. The application also offered a Bob Marley filter, which transformed users' skin tones. This filter was offered alongside similar filters that changes skin tones according to certain ethnicities (these filters were later removed). Regardless of how it is handled, there will always be sensitivity around topics that reference skin colour or ethnicity. Perhaps the filters would have been more successful if user's were presented with a sort-of dial or gradient, where users' could choose their own colour for skin tone instead of predefined categories. This would provide options for both realistic and unrealistic skin tones alike, thus implying neutrality.  
64 ~100 ~https://www.poynter.org/news/how-newsrooms-are-using-machine-learning-make-journalists-lives-easier ~New York Times Research and Development Lab and BBC News Labs ~Their aim is to make the information more "structured" by using tags within HTML or by adding additional metadata to non-textual content such as audio, photo or video. Categorizing content with these tags and metadata pays dividends further down the line, saving journalists time and effort. ~none ~none ~work in progress ~none 
72 ~112 ~https://techcrunch.com/2017/11/08/skype-launches-photo-effects-sticker-suggestions-powered-by-machine-learning/ ~Skpye ~Implementing sticker effects powered by machine learning. ~None. ~The system is capable of features like face detection, including facial expressions, celebrity lookalike generation and various stickers based on this data. Similarly it has object recognition implementation, e.g it can recognise if a dog is in the picture and presents stickers relating to dogs. It can also estimate the users age and determine if they look happy or sad. ~None. ~None. 
232 ~187 ~Article was reported by Will Knight and published by technology review. https://www.technologyreview.com/s/602958/an-ai-ophthalmologist-shows-how-machine-learning-may-transform-medicine/ ~This project was undertaken by Google researchers. ~To design and create an AI Opthalmologist to recognize a common form of eye disease as well as many experts can. ~The algorithm used can look at retinal images and detect diabetic retinopathy as well as a trained opthalmologist can. Google researchers created a training set of 128,000 retinal images classified by at least three opthalmologists. ~Using the deep learning technique, google researchers developed an algorithm to analyze retinal images. Once the algorithm was trained, using the training set of 128,000 retinal images, the researchers tested its performance on 12,000 images and found that it matched or exceeded the performance of experts in identifying the condition and grading its severity. ~Google researchers collaborated with scientists in India, where a clinical trial involving real patients is ongoing. This project involves patients receiving a normal consultation, but their images are also fed into the deep-learning system for comparison. According to Google, results from this trial are not yet ready for publication. ~I think the use of machine learning in medicine is a great idea, especially for certain areas of the world where expertise may be scarce. However, a big challenge for machine learning when it comes to medicine, is to provide convincing evidence that these systems are reliable, as peoples health may be at risk. 
148 ~91 ~University College London (https://www.sciencedaily.com/releases/2017/11/171115091819.htm) ~Dr Parashkev Nachev, Tianbo Xu ~Improve the ability to determine whether a new drug works in the brain ~The research team looked at large-scale data from patients with stroke, extracting the complex anatomical pattern of brain damage caused by the stroke in each patient, creating in the process the largest collection of anatomically registered images of stroke ever assembled. ~Simulated a large-scale meta-analysis of a set of hypothetical drugs, to see if treatment effects of different magnitudes that would have been missed by conventional statistical analysis could be identified with machine learning. ~By illuminating the complex relationship between anatomy and clinical outcome, it enables the detection of therapeutic effects with far greater sensitivity than conventional techniques ~Using machine learning techniques enables researchers to detect the effects of new medicines on the brain with a greater degree of sensitivity than previously used lower level techniques. 
86 ~78 ~http://news.mit.edu/2016/mapping-molecular-neighborhoods-ernest-fraenkel-0705 ~ Ernest Fraenkel ~ biological network modeling to identify new targets for disease.  ~ high-throughput experiments that measure the interactions among all of the molecules during disease ~ using machine learning algorithm to analyse which molecular interaction are specific to the disease ~ it has been used in reducing the time taken by the drugs to reach to customer after proper testing and help in development ~ such research can help reduce the cost of medicne and time taken by them to cure diseases 
99 ~76 ~http://innov8tiv.com/kaspersky-lab-scoops-series-best-class-av-test-awards/ ~Kaspersky Lab ~Create an effective approach to protecting customers against the most sophisticated threats, together with all other forms of cyberattack. ~Not Available ~Mixing Expert Analysts with Big Data/Threat Intelligence and Machine Learning. ~The result was HuMachine, a so called "new benchmark for efficient detection", a protection algorithm based on our global cyber-brain combined with machine learning algorithms and powered by the unequalled human expertise of our security teams, steering our technologies to battle head-on with evolving threats. ~This is a big step in the right direction to help protect everyday users from malware and any other threats. 
149 ~65 ~https://machinelearning.apple.com/2017/11/16/face-detection.html ~Apple ~Apple started using deep learning for face detection in iOS 10. ~they created a large dataset of image tiles of a fixed size corresponding to the smallest valid input to the network such that each tile produces a single output from the network. The training dataset is ideally balanced, so that half of the tiles contain a face (positive class) and the other half do not contain a face (negative class). For each positive tile, we provide the true location (x, y, w, h) of the face.  ~they trained the network to optimize the multitask objective described previously. Once trained, the network is able to predict whether a tile contains a face, and if so, it also provides the coordinates and scale of the face in the tile ~Face recognition on the new iphone  ~This is an incrdible form of security for a mobile phone. 
704 ~260 ~Written by Time Fernholz https://qz.com/915702/the-spacex-falcon-9-rocket-you-see-landing-on-earth-is-really-a-sophisticated-flying-robot/ A paper by Lars Blackmore, principal rocket landing engineer at Space X http://web.mit.edu/larsb/www/nae_bridge_2016.pdf ~The agent is Space X. A privately owned Space company who has Elon Musk as it's CEO. It is not limited to this but most of the paper is discussing the Falcon 9 Rocket so that can also be considered the agent. ~ To improve the uncertainty involved when landing a spacecraft. In the past NASAs Pathfinder that was sent to Mars, had an uncertainty of 150 kilometers around its final landing destination. "In 1997, when NASA sent a rover called Mars Pathfinder to the red planet, it was expected to land within an ellipse 150 kilometers across its major axis, which is not exactly what you want to hear if you're scientist with a specific destination in mind." The main reason for this was because of the use of parachutes to come slow down the craft prior to landing. The goal of Space X is to use rocket thrust to control its descent as well as teaching the rocket to fly itself down. "But once the physics are mastered of maneuvering a rocket-powered spacecraft in for landing, the rocket still needs to be taught to fly itself down."  ~The vehicle uses masses of data. The majority of which is related to the crafts orientation, location, speed, amount of fuel, landing target, destination atmosphere and many other important factors used to succesfully land the F9R. It must compute a divert trajectory and it is imperative that it does this in a fraction of a second. "The vehicle must compute  a  divert  trajectory  from  its  current  location  to  the target, ending at rest and in a good orientation for landing without exceeding the capabilities of the hardware. The  computation  must  be  done  autonomously,  in  a  fraction of a second. Failure to find a feasible solution in time will crash the spacecraft into the ground. Failure to find the optimal solution may use up the available propellant, with the same result." ~"SpaceX uses CVXGEN (Mattingley and  Boyd  2012)  to  generate  customized  flight  code, which  enables  very  highspeed onboard convex optimization." Blackmore and his colleagues are responsible for developing one of the first algorithms to calculate in real time what is needed to land the craft in three dimensions. "He and his colleagues developed one of the first algorithms to do this in three dimensions in that 2009 paper on Mars landings, receiving a patent on their ideas in 2013.The solution involves solving a 'convex optimization problem' a common challenge in modern machine learning". "At SpaceX, Blackmore and his team have updated the landing algorithms, using software developed by Stanford computer scientists'to generate customized flight code, which enables very high speed onboard convex optimization.' As the rocket reacts to changes in the environment that alter its course-known as "dispersions" - the on-board computers recalculate its trajectory to ensure that it will still be 99% sure to land within its target." ~The results have been very successful, in the beginning the first couple of attempts were unsuccessful but they have improved greatly. "So far, it has: SpaceX has landed eight boosters since its first successful attempt in December 2015, including its last four flights. The three failures to land in that time period were caused by hardware issues, not a failure to navigate successfully to the landing area. SpaceX executives are reluctant to say they now expect landings to succeed, preferring to keep their focus on the primary mission of launching cargo for clients, but it's clear that reliability is improving." ~I am genuinely very excited by this. Space X and other companies like Blue Origin are putting the work in place for Humanity to become an interplanetary species. The work they do is facisinating at Space X in particular is fascinating and they seem to be solving major industry problems with high level concepts that can be made seem simple. i.e. reusing space crafts. The future is bright while we have great minds like these providing us with exciting technology enhancements.  
22 ~139 ~https://www.forbes.com/sites/bernardmarr/2017/10/24/how-ai-and-machine-learning-are-used-to-transform-the-insurance-industry/#7d4ae1a913a1 ~Bernard Marr ~Discuss the use of AI in insurance claims ~Consumer survey ~None ~AI helps automate and improve insurance systems ~None 
88 ~728 ~SOURCEChemical Reaction Predictor. http://news.mit.edu/2017/computer-system-predicts-products-chemical-reactions-0627AGENTLarry HardestyGOALA computer system that can instruct the cheapest and easiest way to mass produce an organic chemical.DATAResults and observations from the 100s of sequences of chemical reactions.METHODSApplying machine learning algorithms on data learned from chemical reactions to predict the best possible method to mass produce an organic chemical.RESULTSThis approach saves a lot of the time spent in designing the industrial processes for drug manufacturing.COMMENTSFinding the best methods saves time and money. With time and data; this system could become immensely reliable for drug industries.  
85 ~72 ~https://www.digitaltrends.com/cool-tech/vicarious-ai-research-captchas/ ~Vicarious AI ~Wanted to create a machine learning algorithm (dubbed Recursive Cortical Network) that mimics the human brain in order to break text-based CAPTCHAs.  ~A large collection of training images depicting different characters. ~They downloaded a large amount of CAPTCHA images from various companies such as google, paypal, yahoo.. as well as sample fonts and fed these images into the RCN ~The algorithm was effective in breaking text-based CAPTCHAs with minimal training, it was able to solve googles reCAPTCHAs with 66.6% accuracy ~
399 ~202 ~technologynetworks.com Machine Learning: Helping Determine How a Drug Affects the Brain https://www.technologynetworks.com/tn/news/machine-learning-helping-determine-how-a-drug-affects-the-brain-294252 ~University College London ~To better determine the affects on the brain of the treatment of a certain drug in comparison with conventional statistical models, in particular the effect of a certain drug in brain recovery in stroke patients   ~Large-scale data from patients with stroke, extracting the complex anatomical pattern of brain damage caused by the stroke in each patient. Each entry in the data set featured thousands of variables in order to better model the complex relationships that exist between different elements of the brain. The data collected was at high anatomical resolution. This was viewed as advantageous when compared to current markers investigated for this task, which research felt ignores many crucial variables in determining the impact of a given drug. ~The method employed was a the building of a transductive linear support vector machine classifier with k-fold cross validation to test accuracy. A support vector machine works with numerical data of n dimensions (a dimension for each feature of the data). It separates the data using a line, plane or hyperplane of n-1 dimensions and in this way classifies the data into categories. The parameters of the separating plane are adjusted at each iteration during supervised learning to minimise the classification error. K-fold cross validation involves dividing a test set into k disjoint subsets. At each iteration, train the model over k-1 of these subsets and then test the model on the remaining subset. Change the excluded subset at each iteration.   ~The researchers were able to determine with greater sensitivity the impact of a certain drug on the brain. For example, in the case of lesion reduction, conventional models would need to see a reduction in lesion size of over 78% for the effect of a drug to be detected. In the case of the SVM implemented by researchers, it was successful in determining a positive impact of a drug with just 55% volume shrinkage of a lesion.  ~Researchers feel that machine learning models are far more effective in modeling the brain when compared to statistical models as such models will never come close to replicating the complication of the human brain. For this reason, many results observed using statistical methods when tested on more simplistic animal brains do not hold when the same tests are applied to human brains.  
155 ~117 ~https://www.purdue.edu/newsroom/releases/2017/Q4/system-uses-deep-learning-to-detect-cracks-in-nuclear-reactors.html ~Purdue’s Lyles School of Civil Engineering ~A system to detect cracks in nuclear reactors ~videos of 20 underwater specimens representing internal nuclear power plant components were collected. ~ Samples were scanned at 30 frames per second, and the convolutional neural network examined each frame for cracks.As the data-fusion algorithm observes a crack from one frame to the next it is able to account for changing configurations due to the moving camera, pinpointing the location of the crack. The algorithm mimics the ability of human vision to scrutinize cracks from different angles, which is important because some cracks are obscured by the play of light and shadow ~The approach achieves a 98.3 percent success rate, which is significantly higher than other state-of-the-art approaches ~The Purdue research team also is using deep learning to detect corrosion in photographs of metal surfaces, a technology that might be used to inspect structures such as light poles and bridges. 
348 ~61 ~ http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=7894238 ~ University of California, Los Angeles, CA, USA ~ Predict student performance (final GPA of core courses under consideration, not yet taken, at end of each term) at degree level based on ongoing academic records and student backgrounds with a view to intervening to ensure success towards graduation, on time, and therefore making college less costly.  ~ 1169 undergrad students of Mechanical Engineering (ME) and Arospace Engineering (AE) who completed their course, and excluding those who transferred into the course for whom there was no data. Data for individual students included their GPA and SAT scores from high school(backgrounds) as static features unchanging as student advances through the program, the courses, lectures and labs taken by each, per academic quarter. Letter grade courses were used. Pass/fail courses were excluded. The correlation between SAT scores and their final GPA was referenced. Dept recommended core courses were observed, elective courses were ignored as they tended to be vastly different among the students. ~ The team referenced models used to make course recommendations to students based on best sequence of courses to be taken. Courses were chosen that were relevant to the course results to be predicted, to reduce noise & complexity. This selection was done utilising known pre-requisite courses, combined with courses in the same relevance cluster. They used supervised learning algorithms available to them. (They didn't create their own). Base predictors were constructed using Linear regression, logistic regression, support vector machines, artificial neural networks etc. random forest and k-Nearest Neighbours were the Machine Learning methods used. Clustering was used to get the correlation between courses. ~ As time passed the predictions improved as more data was added. Random forest proved to be the best method used, while KNN performed the worst. The project was considered to be successful, in that the results were found to be useful for both recommending options to students and in identifying cases where interventions might be useful. ~A great deal of information is recorded in the article and presented in graphs which should be read by those interested in learning more. 
96 ~105 ~https://www.newscientist.com/article/2110681-fruity-or-fermented-algorithm-predicts-how-molecules-smell/ ~Andreas Keller, Leslie Vosshall ~Their algorithm tries to guess molecule's smell from it's structure. ~They had 49 volunteers who will rate the odour of 476 chemicals and asked them to match those smells with 19 descriptors. METHOD They released these datas and invited everybody who can develop a logical algorithm. ~Problem was human nature is not that reliable. Because the same volunteers when they get into the test again they were not rating the same odour with the first one. ~Sometimes as we can see in this example, the limits are human nature and behaviours. 
167 ~128 ~Online journal "Artificial Intelligent in Video Games: A Unified Framework" https://www.hindawi.com/journals/ijcgt/2015/271296/ ~Firas Safadi, Raphael Fonteneau, and Damien Ernst ~To understand, express, and demonstrate the pressing need for better artifical intelligence (AI) in video games through a unified framework for conceptual AI development. ~Open-source video game Raven, and StarCraft were used to train and test data. Data consisted of various elements like positioning, targeting, etc. ~Used mainly a C++ open-source framework to write code. Authors developed their own algorithm to improve AI interactions. ~The targeting AI developed by the authors were successful in both Raven and StarCraft. The AI performed better during the testing than the original AI of the games. The authors concluded a similarity between the original and the modified AI of the games, and therefore indicated that a unified framework would be possible in the future. ~Video game AI has been the center of discussion for a long time. I think this may be an example of reinforced learning that I would like to learn. 
96 ~44 ~https://www.deepinstinct.com/#/how-we-do-it ~Deep Instinct ~To be the first company to apply deep learning to the world of cybersecurity. ~Datasets are continuously collected from diverse sources of endpoints and mobile. ~Deep Instinct extracts millions of features per item in a weightless manner, without human involvement. This automatic process enables the identification of first-seen cyber threats. The outcome of this innovative process is the unique ability to identify unknown suspicious behavior. ~The end-result of this innovative process is the ability to identify and prevent even the most evasive cyber-attacks, from any source, on any device (mobile and endpoints). ~
27 ~129 ~https://www.seeker.com/tech/artificial-intelligence/brain-imaging-technology-uses-machine-learning-to-identify-suicidal-thoughts ~Glenn McDonald ~Discuss the use of machine learning to detect suicidal thoughts ~Survey ~None ~AI system could identify suicidal individuals nine out of ten times ~None 
175 ~72 ~Nvidia. https://blogs.nvidia.com/blog/2016/04/11/predict-heart-failure/ ~Shutter Health. Researchers: Edward Choi, Jimeng Sun: Georgia Institute of technology Atlanta.  Andy Schuetz, Walter Stewart: Shutter Health, Walnut Creek California ~To find an improved method for predicting heart attacks ~The data was obtained from the patient records at Sutter Health.  32,784 records were used, 3,884 were those for patients who had heart attacks within an 18 month period and the remainder were used as a control group.  Factors to be considered included diagnosis, medication, procedure codes and summaries of doctor visits. ~The team used Recurrent neural networks and gated recurrent units to find relations between the different factors over time (diagnosis, medication etc).  They also used regularized logistic regression, k,-nearest neighbor, SVM and multilayer perceptrons and used these to compare the prefomance of their model against. ~The GRU model was found to achieve the best results achieving an area under the curve of 0.883. ~The researchers believe that the approach they have taken could be used to predict any disease, not just the risk of a heart attack. 
127 ~123 ~Article on Futurism.com by Dom Galeon https://futurism.com/machine-learning-is-aiding-in-the-fight-against-mental-illness/ ~A team of researchers from several institutions including Carnegie Mellon University and Harvard University. ~To develop a machine learning algorithm trained to understand neural representations of suicidal behaviour. ~The researchers tested their technique in 17 patients with suicidal ideation, and in 17 more that served as control. ~They looked for suicidal brain patterns by watching how the patients' brains reacted when they were presented with six keywords: death, cruelty, trouble, carefree, good, and praise. ~The algorithm was able to accurately identify 15 out of the 17 patients with suicidal ideation, and 16 out of 17 in the control group, for an overall accuracy of 91 percent. ~The results of their study have been published in the journal 'Nature Human Behaviour'. 
47 ~26 ~ https://www.visualdx.com ~ Visual DX ~ a smartphone app that can take images of your skin and accurately on spot perform a full diagnoses  ~ a comprehensive image library of skin conditions ~ un disclosed  ~ accurate on the spot diagnosis and treatment options ~
155 ~94 ~DZone / AI zone https://dzone.com/articles/how-machine-learning-is-helping-drive-cloud-adopti ~John Pollock ~How Machine learning is helping drive cloud adoption ~No dtata mentioned ~The internet of things is big on connecting machines so that they can communicate with one another and exchange data. Machine learning helps to drive these types of interactions, and using the cloud makes it even easier for machines to exchange data with one another as there will be an easy way to make those connections. ~Machine learning and cloud services make an excellent combination, as many cloud services make it easy to provision the resources needed for the collection, storage, and retrieval of large amounts of data. The biggest reasons this works is that such cloud services offer both flexibility and scalability. ~Since everything is on the cloud these days and people rely on it so much for storage and things, machine learning is going to be a big part in anything involved with the cloud. 
101 ~137 ~https://thenextweb.com/artificial-intelligence/2017/10/20/googles-deepmind-achieves-machine-learning-breakthroughs-at-a-terrifying-pace/ ~Google ~To advance the "self learning" of machines. ~Rules of the game Go, and matches of Go ~A machine called DeepMind powers computers called Alpha Go and Alpha Go Zero, the latter which has 4 "AI processors" and the former, which has 48. Both iterations of Alpha Go played against itself. ~Alpha Go Zero defeated Alpha Go in a match of Go. ~It is very interesting the difference in AI processors between the two iterations of Alpha Go. The fact that Alpha Go Zero only knew the rules and still won, demonstrates the potential that this field could amount to. 
28 ~106 ~https://www.newscientist.com/article/2151268-an-ai-has-learned-how-to-pick-a-single-voice-out-of-a-crowd/ ~Mitsubishi Electric Research Laboratory in Cambridge, Massachusetts ~An AI using machine learning to separate the voices of multiple speakers in real time. ~none ~deep clustering ~none ~none 
294 ~224 ~TechCrunch article on Google's "ChatBase": https://techcrunch.com/2017/11/16/googles-chatbot-analytics-platform-chatbase-launches-to-public/ ChatBase website: https://chatbase.com/welcome Area120: https://area120.google.com ~ChatBase is a cloud-based analytics service for builders of conversational interfaces, or chatbots. It helps those builders more easily analyse and optimise their bots for better consumer experiences than ever before. ChatBase is free to use, easily integrates with any chatbot platform, and works with any type of bot, voice or text. Built by area 120, a workshop home to Google's experimental products.  ~The way ChatBase collects data is that it integrates with voice, text or messaging apps like WhatsApp, and look through users' messages to get the data. METHOD The method is used involves clustering. Instead of looking through log files to find patterns in user messages, the system uses clustering for the user's messages, allowing to answer more requests. An application of this clustering found success in Rakuten's "Viber". "We increase the query volume by 35% [...] by optimising queries with high exit rates". By relying on the machine learning to help with the "required optimisations" saving time to build new features. Ultimately ChatBase can handle poorly handled messages with the use of clustering mainly. Also with the use of of conversations it creates a visualisation of flow to improve conversion rates. ~With the use of clustering, the ChatBase app results in giving the clients a more optimised solution to answering more requests and also with these tools, the users can figure exactly where the problem with the bot by identifying and clustering problems, as well as saving time on bot analysis. The main result being that messages can be handled better and so provide a better service to customers and improve the user experience for both bot builders and its consumers, ultimately increasing sales of the bot builders. ~N/A. 
141 ~86 ~https://www.theverge.com/2017/8/2/16082272/google-mit-retouch-photos-machine-learning ~Researchers from Google and MIT. ~To use machine learning models to improve pictures automatically right after they've been captured. This requires a low computational cost from the model. ~Dataset created by Adobe and MIT consisting of 5,000 original images and a set of five touches made to each image by a photography student. ~Convolutional neural networks were used to process a low resolution version of the image and the output was fed to an affine transformation. Finally a three-dimensional grid was used to produce the changes to be made to the photos.  ~They were able to run the application on a mobile platform with low battery drain and low computational cost. The goodness of the produced photos was not evaluated objectively. ~ According to Google's computational photography leader, they are only beginning to scratch the surface with their work. 
278 ~81 ~ https://www.rtinsights.com/netflix-recommendations-machine-learning-algorithms/ ~ NETFLIX ~ Provide users with a continuing supply of recomendations ~ What each user watches, when each user watches things, the place on the Netflix screen the customer found the video, the recomendations the user didn't choose and the popularity of the videos in teh catalog. ~ Several algorithms were used in all in both supervised: classification and regression, and unsupervised: dimensionality reduction through clustering and compression. A video-to-video similarity algorithm is used to generated the suggestions in one of the homescreen rows. This is a form of collaborative filtering but instead of matching users it matches the videos. The algorithm matches watched videos to similar videos and builds a list from these items. It then constructs a table by analysing the videos the user watched and combinin this with data on videos often watched together. This is combined with a Persionalized video ranker algorithm, this algorithm is used to set the order of the videos as they appear in the rows. The closer a video is to the first position increases it's likelyhood of being played however this system is best when also combined with unpersonalized popularity data. ~ Netflix manages to provide highly accurate and relevant recomendations to the user. The video-to-video system is reasonably efficient as most of the computation can be done before and only 1 user must be calculated each time as the table of videos watched together is pre-existing. ~ Quite a good example of strong machine learning as Netflix's entire business model relies on users continually watching more and more videos and seeing as Netflix is incredibly successful we can say that these algorithms definitley work very well. 
307 ~116 ~https://techcrunch.com/2017/11/16/algorithmia-now-helps-businesses-manage-and-deploy-their-machine-learning-models/ ~Algorithmia is the agent who is doing the learning in this story. Algorithmia started out at as an online marketplace for algorithms, most of which were focused on Machine Learning (examples include face detection, sentiment analysis, etc.). With Machine Learning and Artificial Intelligence finding popularity, Algorithmia is launching a new service that will help users to manage and deploy Machine Learning models and to share them easily with others.  GOALS The goal of Algorithmia is to try and turn some of the infrastructure and services it has to run Machine Learning models into a new product.  The co-founder and CTO of Algorithmia stated that usually, R&D spends time collecting and organising data and then building models with that information. He states that Algorithmia instead spent the last five years building an infrastructure that would put the models found by R&D to use.  ~With Algorithmia's new service, data scientist can create models with languages and frameworks of their choice and can host them on the Algorithmia cloud.  ~To create models, Algorithmia offers two services, the Serverless AI Layer and Enterprise AI Layer for hosting the sevice in any public or private cloud. Both versions of the software offer the ability to use got to add the models, to share the models and to handle permissions and authorization. ~According to Katie Gray, Principal of Investments at In-Q-Tel, the CIA's investment arm, Algorithmia has empowered the U.S. government agencies the capability to deply to the AI layer. She stated that Algorithmia delivered a secure, scalable and discoverable platform in which data scientist could focus on problem solving. Algorithmia also announced a $10.5 million funding round led by Google's new AI venture. ~Its impressive how Algorithmia has delivered a platform that uses more than just R&D and actually delivers machine models for data scientists to use. 
206 ~172 ~This article was written by Cheyenne MacDonald and was published by the Daily Mail. http://www.dailymail.co.uk/sciencetech/article-4904298/The-AI-turn-selfie-3D-image.html ~This project was undertaken by researchers at the University of Nottingham and Kingston University. ~To develop an AI that can create a 3D face reconstruction using a single 2D image. ~The researchers trained a neural network on a dataset of both 60,000 2D images and 3D facial models or scans. ~The researchers used a convolutional neural network (CNN) to create 3D face reconstruction. The network learned how to map a face from pixels to 3D coordinates, and essentially works with any picture of a face. It can predict the coordinates of the 3D vertices based on the given 2D image. These reconstructions can be further improved by incorporating 3D facial landmark localizations. ~The outcome of this is a 3D facial model which can work with arbitrary facial poses and expressions as well as lighting issues. This AI can reconstruct an entire face, even adding in parts that might not have been visible in the 2D photo. ~I think this can be a fun experiment to work and play around with using personal images. I feel it could also open up further possibilities such as how we post pictures on social media.   
257 ~88 ~https://www.havashealthandyou.com/wp-content/uploads/2017/06/3.Predictive_Analytics.pdf ~David A. Kvancz, MS, RPh, FASHP; Marcus N. Sredzinski, PharmD; and Celynda G. Tadlock, PharmD, MBA ~This study focused the power of modern analytics on hereditary angioedema (HAE), a single rare disease, because it exhibits features of diseases associated with high costs: rare, hard to diagnose, progressive, and takes a long time from diagnosis to appropriate treatment. Despite the availability of effective therapies, misdiagnoses and underdiagnosis of HAE result in significant burden to the healthcare system.  ~The population for this analysis was extracted from a database of de-identified patient claims data acquired from Truven Health Analytics (MarketScan claims data). This claims database contains health insurance claims data for more than 170 million unique lives covering 2006 through 2014. A 3-stage process was employed to discover patients with HAE within this database who had not been diagnosed. ~The 3-stage process consisted of: identifying people who have been diagnosed with HAE, identify which features or combination of features are most statistically relevant for differentiating HAE from non-HAE patients, and finally use a predictive model to score undiagnosed patients. This score represented the likelihood that the patient had undiagnosed HAE, and patients were ranked from most likely to least likely to have the condition. No description of the model was provided. ~This study successfully demonstrated the ability of this state-of-the-art predictive analysis to find rare-disease patients in a large and complex database. This information could be valuable to claims managers and employers who may realize savings by helping physicians bring these patients to appropriate treatment sooner.  ~
90 ~86 ~http://www.irishmirror.ie/lifestyle/artificially-intelligent-judge-could-help-9114133 ~Researchers at University College London, the University of Sheffield and the University of Pennsylvania. ~Wanted to use an artificially intelligent judge to help deliver verdicts in human cases. ~Researchers used information from 584 cases relating to Articles 3, 6 and 8 on the European Convention on Human Rights. ~The system was asked to label each case as a violation or non-violation by automatically analysing the case text using a machine learning algorithm. ~The The AI Judge was able to predict the verdict of cases with 79 percent accuracy ~
86 ~85 ~https://futurism.com/machine-learning-is-aiding-in-the-fight-against-mental-illness/ ~Carnegie Mellon University, Harvard University and others. ~To tackle the issue of suicidal thoughts. ~The brain activity of 34 patients who experienced suicidal thougts. ~Machine Learning algorithm and MRI. THe patients were presented with words and looking at reactions in the brain. ~Of the 34 patients, 31 were correctly identified as being in the control or having suicidal thoughts. ~This research presents a promising case in the case against suicide and, although still in testing stages, could be used in the future as a prevention.  
109 ~59 ~http://onlinelibrary.wiley.com/doi/10.1111/sltb.12312/full ~John P. Pestian PhD,STM Research Group ~To identify the thought markers of suicidal people ~379 students were used.  ~Each subject completed the Columbia-Suicide Severity rating scale, the young Mania Rating scale and the Hamilton Rating Scale for Depression. Each student then completed a questionnaire and a small five minute interview was coducted. ~Of the 379 subjects. 130 were suicidal while 126 were not and 123  were controls.Overall, the results show that machine learning algorithms can be trained to automatically identify the suicidal subjects in a group of suicidal, mentally ill, and control subjects. The algorithms could detect a a suicidal subject with 85% accruacy  ~N/A 
57 ~118 ~https://www.inc.com/bill-murphy-jr/love-coffee-science-just-gave-you-a-really-healthy-reason-to-have-another-cup.html ~American Heart Association and the University of Colorado School of Medicine ~To determine the effects of coffee consumption on heart health. ~Primarily used data from the Framingham Heart Study. ~Not given. ~A strong correlation between increased coffee consumption and better heart health. ~The study notes that they detected a strong correlative relationship, not a causative one. 
196 ~153 ~https://www.cv-foundation.org/openaccess/content_cvpr_2014/papers/Kazemi_One_Millisecond_Face_2014_CVPR_paper.pdf By Vahid Kazemi and Josephine Sullivan ~KTH, Royal Institute of Technology Computer Vision and Active Perception Lab Teknikringen 14, Stockholm, Sweden ~To show how an ensemble of regression trees can be used to estimate the face’s landmark positions directly from a sparse subset of pixel intensities, achieving super-realtime performance  with high quality predictions. ~A dataset of many facial photos of human beings. ~This  paper  presents  an  algorithm  to  precisely  estimate the position of facial landmarks  in a computationally efficient way.  Similar to previous works:  X. Cao, Y. Wei, F. Wen, and J. Sun.  Face alignment by explicit shape regression.  In CVPR, pages 2887–2894, 2012, P. Dollar, P. Welinder, and P. Perona.  Cascaded pose regression. In CVPR, pages 1078–1085, 2010. The proposed method utilizes a cascade of regressors ~We described how an ensemble of regression trees can be used to regress the location of facial landmarks from a sparse subset of intensity values extracted from an input image. The presented framework is faster in reducing the error compared to previous work and can  also handle partial or uncertain labels.  ~
120 ~95 ~https://www.nytimes.com/2017/11/05/technology/machine-learning-artificial-intelligence-ai.html ~Auto.ML ~To develop machine learning systems that can themselves develop machine learning systems/algorithms. ~Analysises data regarding the performance of various machine learning algorithms and uses it to learn and better its own performance. METHOD: Implementing deep neural networks into this system with the aim of alleviating a lot of the 'heavy lifting' associated with building deep neural networks. Implementing algorithms that analyse the devlopment and performance of other algorithms, in this way the system can learn and output superior algorithms. ~Will significantly accelerate the process of Artificial Intelligence in both online and physical worlds. ~Auto.ML have admitted that it is too early in the development process to be implemented, but it is a case of 'when' and not 'if'. 
40 ~82 ~https://www.engadget.com/2017/05/04/neural-networks-can-help-animate-video-games/ ~Research at University of Edinburgh ~To automatically create animations based on a characters state, geometry of the scene and user input ~Animations and sprites from a large collection of games ~Neural networks ~Results can be seen here: https://www.youtube.com/watch?time_continue=60&v=Ul0Gilv5wvY ~None 
134 ~83 ~https://deepmind.com/applied/deepmind-health/working-nhs/health-research-tomorrow/ ~DeepMind ~To train an AI to learn how to interpret medical test results for themselves and decide which types of treatments are most effective for different patients. ~Head and neck scans at University College London Hospitals NHS Foundation Trust, eye scans at Moorfields Eye Hospital NHS Foundation Trust. ~The system is learning how to identify potential issues within these scans, and how to recommend the right course of action to a clinician. As the algorithm processes more scans, it refines its understanding and interpretation of the information. It then provides increasingly useful feedback, and segmentation, of the data for the clinicians to use for better diagnoses and treatment. ~DeepMind are still in the early stages of AI research in health, and their work is based on deep collaboration with their clinical partners. ~
57 ~468 ~Source:http://news.mit.edu/2017/how-neural-networks-think-0908 AGENT:Larry Hardesty GOAL:understanding nuerul networks that are trained to undertand natural language DATA:no data provided Methods:implementing systems used for object recognision and some intermediate, compact digital representation of the sentence and then try to re-expand it into its original form Results:it was able to translate some of the sentenses that wore given in natural language Comments: 
89 ~75 ~http://news.mit.edu/2015/startup-dmetrics-data-social-media-healtcare-1222 ~Rob Matheson ~Detecting consumer decisions within messy data ~In health care, there’s this gigantic world of unstructured data that needs to be translated into useable information ~Vast stores of messy data, the software reveals insights into consumer decisions ~dMetrics aims to bring its software to more sectors than health care, politics, and consumer finance, with aims of empowering everyone with data. ~The plan was to combine machine learning with natural language processing to decode mountains of unstructured data and provide pertinent information, about anything, to anyone who wanted. 
472 ~115 ~By Aaron S. Jackson, Adrian Bulat, Vasileios Argyriou, Georgios Tzimiropoulos https://arxiv.org/pdf/1703.07834.pdf ~The University of Nottingham, UK and Kingston University, UK  ~Given a single 2D picture of ones face, a 3D enviroment is to be created that can be explored. In other words -> It should be able to reconstruct the whole 3D facial geometry (including the non-visible parts of the face) bypassing the construction (during the training) and fitting (during testing) of a 3D Morphable Model. ~the method requires an appropriate dataset consisting of 2D images and 3D facial scans. As the target is to apply the method on completely unconstrained images from the web, the dataset:  P. Paysan, R. Knothe, B. Amberg, S. Romdhani, and T. Vetter. A 3d face model for pose and illumination invariant face recognition. In AVSS, 2009 was chosen. The dataset of X. Zhu, Z. Lei, X. Liu, H. Shi, and S. Z. Li. Face alignment across large poses: A 3d solution. 2016 was chosen for forming the training and test sets.  ~3DMM - 3D Morphable Model In the 3DMM, the most popular approach for estimating the full 3D facial structure from a single image, training includes an iterative flow procedure for dense image correspondence which is prone to failure.  The work of I. Kemelmacher-Shlizerman and R. Basri. 3d face reconstruction from a single image using a single reference face shape. IEEE TPAMI, 33(2):394–405, 2011, a popular approach for 2.5D reconstruction from a single image, formulates and solves a carefully initialised (for frontal images only) non-convex optimization problem for recovering the lighting, depth, and albedo in an alternating manner where each of the sub-problems is a difficult optimization problem per se. The method of J. Roth, Y. Tong, and X. Liu. Adaptive 3d face reconstruction from unconstrained photo collections. In CVPR, 2016. that produces the average (neutral) 3D face from a collection of personal photos, firstly performs landmark detection, then fits a 3DMM using a sparse set of points, then solves an optimization problem similar to the one in I. Kemelmacher-Shlizerman and S. M. Seitz. Face reconstruction in the wild. In ICCV, 2011., then performs surface normal estimation as in I. Kemelmacher-Shlizerman and  S. M. Seitz. Face reconstruction in the wild. In ICCV, 2011; and finally performs surface reconstruction  by solving another energy minimisation problem. ~Results show that the proposed networks performed well for the whole spectrum of facial pose, and can deal with facial expressions as well as occlusions. Also compared the performance of our networks  against that of recent state-of-the-art methods based on 3DMM fitting reporting large performance  improvement on three different datasets. Future work may include improving detail and establishing a fixed correspondence from the isosurface of the mesh. ~The deep learning of a mapping from pixels to 3D coordinates using a Convolutional Neural Network (CNN). 
99 ~74 ~http://news.mit.edu/2017/using-machine-learning-improve-patient-care-0821 ~CSAIL ~To predict patient outcomes in hospitals; to determine what kinds of treatments are needed for different symptoms. ~Large amounts of intensive-care-unit data, from vitals and labs to notes and demographics. ~The system uses deep learning to make real-time predictions, learning from past ICU cases to make suggestions for critical care, while also explaining the reasoning behind these decisions. ~Patients relieved of ailments quicker and diagnosed with more accuracy ~This technology is still in early development. Along with quick diagnosis this technology could potentially be an aid for doctors in the ICU, which is a high-stress, high-demand environment. 
72 ~85 ~https://futurism.com/machine-learning-is-aiding-in-the-fight-against-mental-illness/ ~Carnegie Mellon University and Harvard University ~Create a machine learning algorithm that can identify if someone has sucidal tendencies from an MRI scan. ~17 patients with suicidal ideation and 17 people that served as a control ~Used a machine learning algorithm to analyse the MRI scans of the people in the hope to find abnormalities ~Found the article interesting but I think there is not enough data to prove the point 
120 ~108 ~https://www.newscientist.com/article/2143498-deepmind-ai-teaches-itself-about-the-world-by-watching-videos/ ~Relja Arandjelović ~Removing the dependence of machines on people in the machine learning. ~The system was trained on 60 million still-audio pairs taken from 400.000 videos. METHOD He created his algorithm into three networks.One is for recognising images, second one is for audio and the third one is trying to form a relationship between still-images and audios. He calls co-learning to this system. ~This algorithm can classify an audio clip 80 percent of the time. ~I do not want to look like a conspiracy theorists but unlike humans, machines won't get tired or bored or lsoe focus on learning.So what is important here for me, what are we going to be able to use those informations for our interests. 
96 ~72 ~https://www.digitaltrends.com/cool-tech/vicarious-ai-research-captchas/ ~AI company Vicarious. ~To develope a program that can solve CAPTCHAs meant to detect bots. ~CAPTCHA datasets from Google, Yahoo, Paypal and data from BotDetect system. They also used the MNIST dataset with added occlusions and noise. ~They created a model called Recursive Cortical Network (RCN), which is an object-based model that assumes factorization of contours and surfaces, and objects and background. ~The system was able to bypass BotDetect system with an accuracy of 57 percent with less training than conventional neural networks. ~CAPTCHAs are still ubiquitous online despite their diminishing effectiveness in bot detection. 
48 ~123 ~ https://siliconangle.com/blog/2017/11/13/microsoft-applies-machine-learning-deliver-neural-fuzzing-vulnerability-testing/ ~ Microsoft  ~ to use machine learning and neural networks to detect security flaws in existing systems using past experiences  ~ examples include previous overlooked issues and security leaks from Microsoft itself  ~ Neural Fuzzing  ~ yet to be disclosed  ~
90 ~96 ~Apple Machine Learning Journal https://machinelearning.apple.com/2017/11/16/face-detection.html ~Apple engineers. ~Using deep learning to improve the iPhones facial detection utilizing only onn device information. ~iPhone owners photo library. ~Deep Learning using the OverFeat approach. This approach utilises deep convolutional networks to scan images. ~Apple successfully managed to create a deep learning facial recognition algorithm that was able to correctly identify if a face was present in an image tile while only running on device.  ~This is impressive due to the resource limitations posed by running the entire algorithm on a mobile phone. 
408 ~113 ~https://www.technologyreview.com/s/537366/the-machine-vision-algorithm-beating-art-historians-at-their-own-game/ ~The agents here are Babak Saleh and Ahmed Elgammal, for their paper, "Large-scale Classification of Fine-Art Paintings: Learning The Right Metric on The Right Feature" ~The goal of this research is to use machine learning methods to train algorithms to recognize the arts and style of a fine-art painting. The machine should be able to make judgements related to aestehtics on a semantic level. ~Elgammal and Saleh used the 'Wikiart paintings' dataset set, which comprises 81,449 fine-art paintings from 1,119 artists across fifteen centuries. Within this are contained 27 different styles, and 45 genres.  ~Three methodologies were used: Metric Learning, Feature Fusion, and Metric-Fusion. Metric Learning finds some pair-wise real-value function which is non-negative, symmetric, obeys the triangle inequality, and returns zero if and only if x and x-prime are the same point. Neighborhood Component Analysis, Large Margin Nearest Neighbor, Boost Metric, Information Theory Metric Learning, and Metric Learning for Kernel Regression are also used. These algorithms work together to highlight certain features in a given art piece - these can be low-level features such as overall colour, and more advanced features such as image identity - maybe a horse or a table in the piece. This produces a description of the painting with 400 different dimensions. ~When the method is complete, it is tested on a set of paintings it has not yet seen. This can accurately identify the artist in voer 60% of the paintings, and the style in 45% of them. It also some produces some perhaps unexpected results. In one example it has difficulty distinguishing between the works of Pissarro and Monet. However, these two artists were close friends with many shared experiences, so would have produced similar pieces at a level of detail the human eye may not pick up on immediately. It tends to find links between some styles such as expressionism and fauvism - the latter is a subgenre of expressionism. Links such as these are known by art historians but perhaps not to most pedestarian art consumers - however the algorithms picked up on it.  ~ This an extraordinary use of Machine Learning methods and one not many could think of off-hand. Art seems to be a big draw for Machine Learning research, another project that comes to mind is one of Tubingen, Germany, where researchers have developed an algorithm that morphs any given photo to resemble the style of one of the 'great masters' of art. 
301 ~73 ~https://qz.com/495614/computers-can-now-paint-like-van-gogh-and-picasso/ ~No agent was listed in the study, however it was a deep nerual network similar to Google's Deep Dream system. ~To use neural representations to separate and recombine content and style of arbitrary images, providing a neural algorithm for the creation of artistic images. ~The Dataset for this study was based on a large selection of well known paintings, these paintings were photographed, in high resolution to ensure accuracy, the dataset consisted of serveral sub sets of paintings in a certain style. ~Convolutional Neural Networks - consists of layers of small computational units that process visual information hierarchically in a feed-forward manner. Each layer of units can be understood as a collection of image filters, each of which extracts a certain feature from the input image. Each style of painting was tested indivudially, that was a certain style could be analysed before the agent attempted to recreated a paiting in this style. The researchers taught the system to see how different artists used color, shape, lines, and brushstrokes so that it could reinterpret regular images in the style of those artists. ~The key finding of the paper is that the representations of content and style in the Convolutional Neural Network are separable. That is, you can manipulate both representations independently to produce new, perceptually meaningful images. To demonstrate this finding, they generate images that mix the content and style representation from two different source images. In particular, they could match the content representation of a photograph depicting the Neckarfront in Tubingen, Germany and the style representations of several well-known artworks taken from ¨ different periods of art. ~This study shows how well a machine can interpret a specific artists style, in such a degree that it can create a combination of that style and others if it was instructed to. 
47 ~111 ~https://www.infoworld.com/article/2907877/machine-learning/how-paypal-reduces-fraud-with-machine-learning.html ~PayPal ~Fraud Detection ~PayPal collects large amounts of data about buyers and sellers, including their network information, machine information, and financial data. ~PayPal uses three types of machine learning algorithms: linear, neural network, and deep learning. ~Paypal's deep learning is able filter out fraudulent transactions. ~
183 ~98 ~https://www.newscientist.com/article/2110522-googles-neural-networks-invent-their-own-encryption/ ~ Martin Abadi, David Andersen and researchers at Google Brain, Googles's Deep Learning dept. ~To use machine learning to teach computers how to encrypt messages. ~Plaintext chosen by the researchers to be encrypted. ~Three computers - Alice, Bob and Eve. Alice had to send an encrypted message to Bob who had to decrypt it, while Eve had to eavedrop and attempt to decrypt the message. They used  repeated trials to allow the machines to learn how to come up with their own encryption algorithm and to improve it. ~The machines were slow to learn, but after 15,000 trials Alice and Bob were successful, and Eve could only decrypt half of the encrypted bits, which is the same as pure chance as the bits are either 1 or 0. ~Because the machines had to come up with their own encryption algorithm, it is difficult to tell how secure the algorithm is. At present there aren't many uses for this technology, but as machine learning is improved upon these techniques could grow and could be used to securely encrypt our data against hackers. 
119 ~199 ~Machine-learning tech helps detect illegal drug peddlers on Twitter - http://indianexpress.com/article/technology/science/machine-learning-tech-helps-detect-illegal-drug-peddlers-on-twitter-4907409/ ~Researchers at Twitter. ~Researchers have developed a machine-learning technology that mined microblogging site Twitter to identify users peddling the illegal sale and marketing of prescription opioids online. ~Researchers used Machine learning to isolate tweets related to the marketing of opioids, and web forensic examination to analyse posts that included hyperlinks to external websites. The researchers collected some 619,937 tweets containing the keywords codeine, Percocet, fentanyl, Vicodin, Oxycontin, oxycodone and hydrocodone between June and November 2015. Of these, they found 1,778 posts that were marketing the sale of controlled substances, 90 per cent included hyperlinks to online sites for purchase. ~None were mentioned. ~The technology developed could be u 
106 ~85 ~Apple Machine Learning Journal https://machinelearning.apple.com/2017/07/07/GAN.html ~Apple engineers. ~Improving the Realism of Synthetic Images. ~Sets of synthetic images as well as real images of eyes without labeling. ~Adversarial discriminator network, Generative Adversarial Networks. ~10 subjects chose the correct label 517 times out of 1000 trials, meaning they were not able to reliably distinguish real images from refined synthetic ones. In contrast, when testing on original synthetic images vs real images, we showed 10 real and 10 synthetic images per subject, and the subjects chose correctly 162 times out of 200 trials. ~Allowing for easier generation of realistic synthetic images will allow for easier training of future networks. 
22 ~105 ~http://www.popularmechanics.com/space/deep-space/news/a28752/an-ai-found-dozens-of-gravitational-lenses/ ~Universities of Groningen, Naples, and Bonn. ~Search for gravitational lenses. ~Kilo-Degree Survey. ~Convolutional neural network ~Found 56 gravitational lens candidates. ~
83 ~354 ~https://motherboard.vice.com/en_us/article/8qb5mv/new-machine-learning-program-recognizes-handguns-in-even-low-quality-video https://arxiv.org/pdf/1702.05147.pdf https://www.engadget.com/2012/01/21/nypd-begins-testing-long-distance-gun-detector-as-alternative-to/ http://money.cnn.com/2014/10/23/smallbusiness/schools-gun-technology/?iid=article_sidebar ~A team of computer scientists based at the University of Granada. ~Recognising gun in video. ~ImageNet 1.28million images. + 3000 extra hangun images. ~VGG-16 model finetuned with 3000 handgun images. ~Recognises handguns in even low qualities youtube videos. It does this in under 0.3 seconds. An alarm is raised if more than 5 positives are found in under 0.2 seconds or 27 scenes. ~The solution has two competitors. One based of heat reading and another for recognising gunshots sounds. 
46 ~151 ~http://www.bbc.com/news/technology-40681395 https://www.newscientist.com/article/2141363-ai-suggests-recipe-for-a-dish-just-by-studying-a-photo-of-it/ ~Researchers at the Massachusetts Institute of Technology (MIT). ~To identify recipes for food based solely on a photograph. ~One million photos and one million recipes. ~Neural networks trained on the above photographs. ~The system picked the right recipe 65% of the time. ~No comment. 
88 ~77 ~https://www.engadget.com/2017/11/10/counterfeit-ai-machine-learning-forgery/ ~University of Washington ~Use deep learning to produce fake realistic voice clips and video images. ~using 1,000 videos containing some 46,000 sounds resulting from different objects being poked struck or scraped with a drumstick. ~None specified. ~In the future it can make it too hard for counterfeiters to create accuraet currencys as it will take too much to make it profitable. A staggering amount of people could hardly tell the difference in the clips. people actually chose the fake audio over the real twice as often ~NA 
123 ~151 ~https://www.forbes.com/sites/tomdavenport/2017/11/05/revolutionizing-radiology-with-deep-learning-at-partners-healthcare-and-many-others/#26632fd95e13 ~Center for Clinical Data Science ~goal is to employ machine learning and other artificial intelligence technologies to improve the healthcare delivery system; in particular, a key CCDS objective is to improve the effectiveness of imaging-based diagnosis. ~One of the key resources for more general image recognition projects, for example, is the open source ImageNet database, with over 14 million labeled images.  ~The CCDS is pursuing a variety of machine learning approaches, but the primary technology that it is employing is deep neural networks (also known as deep learning).  ~No results yet as this requires time but the objective is harnessing the power of deep learning for medical image analysis ~This is the capability to make the world a better place 
119 ~135 ~https://futurism.com/machine-learning-is-aiding-in-the-fight-against-mental-illness/ https://www.nature.com/articles/s41562-017-0234-y ~ A team of researchers from Carnegie Mellon University and Harvard University. ~ A machine learning algorithm trained to understand neural representations of suicidal behaviour. ~ 17 patients with suicidal ideation and 17 others that were to serve as a control. ~ Patient's brains were monitored, using a MRI machine, while being presented with 6 key words: "death, cruelty, trouble, carefree, good, praise". ~ Correctly identified 15 out of the 17 patients with suicidal ideation and 16 out of the 17 control for an overall accuracy of 91 percent. ~ While the results of these tests are high, it would be difficult to implement, in a practical sense, as it requires the use of an MRI machine. 
103 ~112 ~https://www.technologynetworks.com/tn/news/scientists-use-machine-learning-to-analyze-language-in-movies-294179 ~ University of Washington ~ The goal is to identify the power and agency of female characters in movies. ~The last 20 years of movies and 21000 different characters were used as data. METHOD  The dialog of the female character was analysed, and their speech was ranked based off how imperative their statements were, i.e. words like instructs would rank highly but words like implores would rank lowly. RESULT They found that while female characters were getting more agency in the plot their dialog still reinforced traditional female stereotypes. Males characters scored higher on both power and agency in all genres.  
28 ~71 ~http://news.mit.edu/2017/robot-learns-to-follow-orders-like-alexa-0830 ~MIT Researchers ~A system that allows robots to understand a wide range of commands that require contextual knowledge about objects and their environments. ~none ~none ~none ~none 
108 ~100 ~Google Research: Acoustic Modeling for Google Home (https://research.google.com/pubs/pub46130.html) ~Bo Li, Tara Sainath, Arun Narayanan, Joe Caroselli, Michiel Bacchiani, Ananya Misra, Izhak Shafran, Hasim Sak, Golan Pundak, Kean Chin, Khe Chai Sim, Ron J. Weiss, Kevin Wilson, Ehsan Variani, Chanwoo Kim, Olivier Siohan, Mitchel Weintraub, Erik McDermott, Rick Rose, Matt Shannon ~Improve Sound quality of connected speaker by applying derverberation with neural network. ~22 millions of English utterances collected from Google Voice search service. Applying artificial noise and reverberation. Work on raw waveforms. ~Logistic Disjunctive Normal Networks and asynchronous stochastic gradient descent optimization. ~Huge improvements on environment with reverb and improvement of 18% compared to the previous system. ~
364 ~204 ~ https://twitter.com/tayandyou?lang=en https://www.cnbc.com/2016/03/30/tay-microsofts-ai-program-is-back-online.html https://gizmodo.com/here-are-the-microsoft-twitter-bot-s-craziest-racist-ra-1766820160 ~ Microsoft Tay was created Microsoft's Technology and Research and Bing teams. ~ Tay was released in Twitter with the handle @TayandYou, on 23 March 2016. The chat-bots current biography reads 'The official account of Tay, Microsoft's A.I. fam from the internet that's got zero chill! The more you talk the smarter Tay gets.' Tay was intended to mimic the language patterns of it's target audience (18 to 24 year old Americans). Thus functionality such as creating memes from users' photographs was entrusted to it. ~ Tay had been modelled after Xiaoice; a chat-bot that had previously been released into the Chinese community through Weibo. Both Xiaoice and Tay learned from interactions through social media, however the integration of the Western chat-box proved less successful. ~ Tay uses machine learning algorithms to learn from the mannerisms of its correspondents. ~ Soon after its release, Tay began to tweet politically incorrect posts that mimicked the tweets that it had been receiving. This resulted in the accounts first suspension, 16 hours after its release . The decision was probably made because of the bad publicity and general uproar in the public realm. Microsoft issued an apology for the inappropriate tweets made by the chat-bot.  Tay was "accidentally" re-released on 30 March, 2016; tweeting yet more politically incorrect posts. ~ Personally I think that Tay was successful in achieving what the chat-bots goal was (to learn from interactions with users in social media). No news is bad news, and however "bad" the publicity was surrounding the endeavours of Tay, it was publicity all the same. What this has really shown is the differences between Eastern and Western cultures (or perhaps it is a display of censorship). There was no claim of Tay being the perfect and proper chat-bot; the claims made by Microsoft were that Tay would become "smarter" the more it is interacted with.  Tay certainly learned how to interact with others through how it itself was interacted with. There was no topic blacklisted by Microsoft's developers. This decision has been commented on as an oversight by members of the public. I ask the question: Was it really an oversight? 
48 ~80 ~http://news.mit.edu/2016/artificial-intelligence-produces-realistic-sounds-0613 ~MIT's Computer Science and Artificial Intelligence Lab ~Generate audio from a silent video feed ~Approx 1000 videos of roughly 46,000 sounds ~Deconstruction of sound and analysis of pitch, loudness etc ~Online study where participants had to choose between real and generated sounds showed a realistic result ~
553 ~289 ~I found the source at this url https://medium.mybridge.co/machine-learning-top-10-articles-for-the-past-month-v-oct-2017-c87211085729. It had a link to the blog by Trapit Bansl, Igor Mordatch, Jakub Pachocki, Ilya Sutskever and Szymon Sidor. https://blog.openai.com/competitive-self-play/ ~The agent here is OpenAI and their "Competitive Self Play" model. ~The goal here is to have various versions of the agent train and learn by playing against eachother competitively and without the need of human designed tasks and environments. The reason they want to move away from the need of human intervention is because "the agents' behaviours will be bounded in complexity by the problems that the human designer can pose for them". By using the "Competitive Self Play" model they can develop the agents much faster and more successfully. " By developing agents through thousands of iterations of matches against successively better versions of themselves, we can create AI systems that successively bootstrap their own performance; we saw a similar phenomenon in our Dota 2 project, where self-play let us create an RL agent that could beat top human players at a solo version of the e-sport."  ~The data they work off is based off a rewards system each agent is acting on. The also act and react according to their opponent's moves. They create more data as they progress, working off new methods learned. The environments the agents operate in are data and how to win particular games is more data they work off. Their data set is constantly evolving as they learn new skills. ~"Each agents neural network policy is independetly trained with 'Proximal Policy Optimization'". They provide agents with rewards for behaviour which encourage the agents to explore new movements in the beginning."Here we took the dense reward defined in previous work for training a humanoid to walk, removed the term for velocity, added the negative L2 distance from the center of the ring and took this as a dense exploration reward for our sumo agents." They then steadily converge the rewards to zero as the agent improves in favour of rewarding them for winning and negatively for losing. They set the agents up in different scenarios/environments, like sumo wrestling, get by other agents, one defend while the other attempts to score and more. Over time, they begin to learn new skills and strategies. These skills can be transferred to other environments that have not been experienced by the agent. For example the skills learned from sumo wrestling are used to remain standing in a new environment where a wind of different forces is applied to the agent.   ~This seems to have been very succesful in teaching agents new skills without the need of a human designer. It is also very positive in the sense that the skills learned can be used in different scenarios/environments. "We demonstrated the development of highly complex skills in simple environments with simple rewards. In future work,  it would be interesting to conduct larger scale experiments in more complex environments that encourage agents to both compete and cooperate with each other." ~One feels this could be extremely beneficial to the growth of deep learning. The way the team demonstrated how the agents can learn so rapidly while using "Competitive Self Play" is very intriguing and no doubt many others will undertake similar paths when designing their machine learning agents in future.  
186 ~110 ~ http://med.stanford.edu/news/all-news/2017/11/algorithm-can-diagnose-pneumonia-better-than-radiologists.html ~Standford Medicine.  ~ To use ML to process chest xrays with a view to diagnosing 14 types of medical conditions as a compliment to the work of radiologists, by removing bias and overcoming some of the problems associated with human perception and reducing errors. The success of the project would contribute towards improved health care delivery. ~ The team started with a public dataset of 112,120 x-rays covering the 14 pathologies was used.  ~ An algorithm that could diagnose these 14 patholgies was available. The team chose to use this dataset and the algorithm in place to focus on indications for pneumonia.  ~Pneumonia diagnosis proved better than that of radiologists. The new ML algorithm developed by the team more accurately diagnosed all 14 patholgies than the previous algorithm and than the diagnosis of their Standford radiologists in terms of pneumonia which is regarded as being especially difficult to diagnose. ~Future potential for the project is to highlight areas in the chest x-rays to radiologists which the algorithm suggests are most likely to suggest pneumonia. thus reducing the numbers of missed diagnosis. 
130 ~75 ~http://www.itweb.co.za/index.php?option=com_content&view=article&id=166475 ~Vodacom  ~When you use your app or the USSD platform, the bundles you see are personalised for you, which makes your propensity to buy much higher. ~Customer Behaviours on the platform ~As you buy it keeps teaching itself and there are thousands of different packages and combinations. The machine is then learning, based on you and your segment, and teaching itself ~The platform seems to be working, as Vodacom's total bundles sold in SA increased 64.8% to 1.1 billion in the six months ended 30 September ~Gauss was really smart but I wonder if the data was actually Gaussian distributed.  He unfortunately did not generate figures that would have made this easier to check, perhaps due to not having a computer and instead doing everything by hand. 
291 ~181 ~Machine Learning: Helping Determine How a Drug Affects the Brain - https://www.technologynetworks.com/tn/news/machine-learning-helping-determine-how-a-drug-affects-the-brain-294252 ~Researchers from University College London were involved in this study. Dr Parashkev Nachev (UCL Institute of Neurology) was lead researcher in this study. The study was funded by Welcome and the National Institute for Health Research University College London Hospitals Biomedical Research Centre. ~Machine learning could improve our ability to determine whether a new drug works in the brain, potentially enabling researchers to detect drug effects that would be missed entirely by conventional statistical tests. ~The dataset used was a large-scale meta-analysis of a set of hypothetical drugs, to see if treatment effects of different magnitudes that would have been missed by conventional statistical analysis could be identified with machine learning. ~Using a support vector machine (SVM) training algorithm researchers built a model that assigns new examples to one category or the other, making it a non-probabilistic binary linear classifier. The image above is represented as a 3D cubic glyph varying in colour and scale, it is transudative linear support vector machine classifier trained to relate the high-dimensional pattern of damage to gaze outcome. ~Current statistical models are too simple. They fail to capture complex biological variations across people, discarding them as mere noise. Researchers looked at large-scale data from patients with stroke, extracting the complex pattern of brain damage caused by the stroke in each patient, creating in the process the largest collection of anatomically registered images of stroke ever assembled. ~It is worth mentioning that machine learning adds a real value, it formalises complex decisions which can be very effective. It is evident from this article simple algorithms such as the SVM algorithm can capture the complex anatomy and high precision in the field of medical science. 
74 ~58 ~http://fortune.com/2015/10/16/how-tesla-autopilot-learns/ ~Tesla ~To bring an efficient, smart, and reliable Auto-pilot to cars. ~Driving dictionary, A data set full of thousands of mules of driving videos, GPS data and more... ~Not Specified ~As seen today, there already are Tesla cars capable of Auto-piloting with a highly reduced crash risk compared to a human driver ~This is where the future is leading, and with this we can see the progression in our technology is incredibly rapid.  
61 ~111 ~http://spacenews.com/with-commercial-satellite-imagery-computer-learns-to-quickly-find-missile-sites-in-china/ ~National Geospatial Intelligence Agency(USA) ~Automation of image analysis tasks which are repetitive and time-consuming. ~Satellite imagery data sets. ~Deep learning neural network. ~80 times more efficient than a human doing visual searches for surface-to-air missile sites. Had a 90 accuracy, the same overall statistical accuracy as human analysts. ~Possible automation of a large chunk of jobs currently done by humans. 
144 ~31 ~https://serenatadeamor.org/en/ ~Irio Musskopf ~Analyse reimbursement claims from congresspeople in Brazil and identifying the probability of it being illegal. ~Receipts submitted by the politicians. ~The project doesn't go into too much detail about the learning methods, only that k-means is used to cluster the data. ~The project is still onogoing and, at the time of writing, there are 8276 suspicious reimbursements found by 735 different congresspeople. After a claim is deemed suspicious, there is a human element of checking wheter the claim is tru or not. ~This is a really interesting use of machine learning. What makes it more interesting is that the project is open source and anyone can contribute to it. Unfortunately the process cannot be fully automatised due to the legal element, but if it was possible to do so, it would raise some interesting questions. Can machines pass judgment on people?  
98 ~45 ~http://time.com/4928010/diagnose-cancer-pen/ ~Researchers at the University of Texas at Austin. ~Wanted to distinguish between cancerous tissue and healthy tissue, during tumour removal. ~Analysing small molecules produced by human cells and using pattern identification to identify unique sets of cells associated with cancer. ~The MasSpec pen produces a small drop of water that extracts molecules from a persons cells during surgery. Through machine learning they are able to determine what molecular fingerprint is normal and what is cancer. ~The handheld device is able to analyse human tissue samples for cancer with 96 percent accuracy. ~The pen is not yet DFA-approved 
115 ~155 ~https://www.scientificamerican.com/article/worlds-most-powerful-particle-collider-taps-ai-to-expose-hack-attacks/ and https://arxiv.org/pdf/1704.06193.pdf ~Andres Gomez et al. (see https://arxiv.org/pdf/1704.06193.pdf) ~Increase security, specifically with respect to the ALICE (A Large Ion Collider Experiment) Pb-Pb detector's production environment, specifically, via intrusion prevention (no ML involved) and detection (ML involved). ~All data generated during job runs (log entries, traces, syscalls, etc.) ~Further reading into the arxiv paper revealed that the machine learning techniques are 'proposed' for job metric analysis, essentially allowing them to detect abnormal situations by learning what a normal situation looks like (binary classification). They suppose SVMs (support vector machines) may be useful due to their successful application in other security applications. ~No results as of yet, since the paper merely promises future ML work. ~
42 ~96 ~Source:https://www.digitalhealth.net/2017/11/machine-learning-delivers-insight-stroke-patients/ ~GOAL:To see what effects drugs have on stroke victims ~Methods:Applied machine leaning algorithms to mri and ct sans of stoke patients  Results:Treated each patients seperatly and gave the presense or absense of damage in patients who had a stroke Comments: 
381 ~171 ~"Uber Data to Determine The Best Food Places In New York City" - https://www.forbes.com/sites/lutzfinger/2017/10/19/uber-data-determine-food-places-new-york/#551d1cc969b7 ~This project was collaboration with Katherine, Swapnil Gugani, Tianyi Tang, and Wu Peng. ~The goal was using public Uber data to see what is the best real estate, and how can a food truck best optimize their location to maximize sales. ~This dataset was taken from using public Uber dataset, pickup and drop off points that occurred over a 3-month time window within Manhattan were observed. This data roughly reflects high traffic locations in NY and thus can be a primer for food truck locations. The data set comes in the form of spatial coordinates. ~The machine learning algorithm used was a k-means clustering algorithm, it was used to pinpoint cluster centres within the traffic data. The underlying idea is that a cluster centre generated from this dataset would generate spots on the map that minimize distances between pickup points, indicating locations with ideal points to set up food carts to access the highest number of customers. ~It is interesting to notice that in figure 1 there are significant differences between the cluster centres and the top-ranked points at different times. Main centres of pickups are evident, especially for smaller, fully mobile carts this kind of information could help tell operators where to go to take advantage of Uber customers. There were some forthcomings using the k-means clustering algorithm, the distance is Euclidean and not along the actual roads, so it might be that centres seems close but in reality, it is not. Moreover, this assumes that Uber users are good food truck customers. To test the hypothesis that there is a relationship between Uber pickups and food truck locations, the data was triangulated with food truck location data scraped from Yelp. The article continues talking about the correlation between food truck locations and Uber pickups and how indicating locations with ideal points to set up food carts could lead to accessing the highest number of customers. ~An analysis such as this is very powerful and effective, it shows how app data from sources such as Uber could be used to inform seemingly unrelated businesses, showing the potential to incorporate external data sources in addition to internal ones. K-means clustering algorithm is definitely effective is such cases. 
209 ~93 ~Example 5:Understanding deep features with computer-generated imagery http://imagine.enpc.fr/aubrym/projects/features_analysis/texts/understanding_deep_features_with_CG.pdf ~ Mathieu Aubry Ecole des Ponts  Bryan C. Russell Adobe Research   ~Approach to analyzes the variation of features generated by convolutional neural networks in scene factors that occur in natural images such as 3D viewpoint, color, and scene lighting configuration  ~Images created by CNN's  AlexNet, Places  and Oxford VGG ~The application begins by rendering a set of stimulus images by varying one or more scene factors. Then they present the stimuli images to a trained CNN as input and record the feature responses for a desired layer. Given the feature responses for the stimuli images they analyzed the principal modes of variation in the feature space via PCA. ~They have introduced a method to qualitatively and quantitatively analyze deep features by varying the network stimuli according to factors of interest. Utilizing large collections of 3D models, they have applied this method to compare the relative importance of different factors and have highlighted the difference in sensitivity between the networks and layers to object style, viewpoint and color. ~Another bedrock project which will be used further to improve neural networks ability to render 3d scenes and allow future CNN's in the to produce better 3d rendering. 
211 ~297 ~https://petapixel.com/2017/05/18/ai-powered-app-helps-colorize-black-white-photos-seconds/ https://petapixel.com/2016/03/31/photoshop-future-may-able-auto-colorize-bw-photo/ https://richzhang.github.io/colorization/ https://dev.to/developius/colourising-video-with-serverless-machine-learning-c8a ~Originated by Richard Zhang and built on by a team of researchers at UC Berkeley. Utilized by Photoshop and expected to be incorporated in the next release. ~To automatically colorize black and white photos ~It is easy to convert colored images to grayscale. Hence, millions of colored images were used as a training set. ~A feed-forward convoluted Neural Network was used. The training set consisted of millions of colorized photos. These were grayscaled and sent in as an input to the FFNN, and trained on the original colorised version as an output. This is a form of supervised machine learning. 8 layers were used in the FNN. This form of deep machine learning is necessary. The first layer determines an overall estimate of the scene and its appropriate colors. As the Neural net gets deeper more specific colors are applied to individual objects. ~20 percent of the photo pairs actually fooled the humans. This means the colorized versions were guessed as the real color photos. This fool rate is much higher than prior research in this area. ~As this becomes more efficient, the algorithm can be applied to frames to create colorized videos. At this moment it costs about 2000 euro per minute to colorize a video. 
148 ~103 ~http://www.zdnet.com/article/how-machine-learning-is-helping-virgin-boost-its-frequent-flyer-business/ ~Rutgers University and the Atelier for Restoration & Research of Paintings in the Netherlands. ~Wanted to use AI to to detect forgeries. ~They used line drawings by Picasso, Matisse, Modigliani and other famous artists. Also drawings by other artists in the same style to test spotting fakes. ~The system broke almost 300 line drawings into 80,000 individual strokes and a deep recurrent neural network (RNN) learned what features in the strokes were important to identify the artist. They also trained a machine-learning algorithm to look for specific features, like the shape of a line in a stroke which allowed them to detect forgeries. ~Using the algorithms mentioned they were able to correctly identify artists around 80 percent of the time. Using the drawings created by artists in the style as the pieces in the dataset, the system was able to correctly identify forgeries in every instance. ~
147 ~123 ~Article by Chris Raphael on RTInsights.com https://www.rtinsights.com/netflix-recommendations-machine-learning-algorithms/ ~Netflix ~The goal is to determine what shows/movies Netflix should recommend to their each of their users, based on each users' viewing habits. ~Netflix has access to data about what each of their members watch, when they watch, the place in Netflix the user found the video, recommendations the customer didn't choose, and the popularity of videos in the catalog. ~All of the data get fed into several algoritms powered by statistical and machine-learning techniques. Approaches use both supervised (classification, regression) and unsupervised (dimensionality reduction) approaches. ~Over the years, the recommendation system has decreased customer churn and saves the company about $1 billion a year. Personalized content also helps find an audience for relatively niche videos which would not be viewed other wise. ~Netflix also use "evidence" algorithms, which focus on what information to show a viewer about a movie. 
459 ~330 ~Macrumours.com - Apple Says 'Hey Siri' Detection Briefly Becomes Extra Sensitive If Your First Try Doesn't Work https://www.macrumors.com/2017/10/18/apple-explains-how-hey-siri-works/ Apple - Hey Siri: An On-device DNN-powered Voice Trigger for Apples Personal Assistant https://machinelearning.apple.com/2017/10/01/hey-siri.html ~Apple ~To recognise when a user says the words, Hey Siri and to respond by opening up the Siri assistant tool on an Apple device.   ~The data in this case is an audio extract provided by the user. A small speech recogniser runs constantly inside an apple device and is constantly listening to user speech and trying to detect if the user has said hey siri or not. The acoustic model is trained on utterances of Hey Siri that were detected before the automatic Siri activation feature was available. Apple found that many users already said Hey Siri upon starting a conversation with Siri in the traditional way (using buttons) and therefore they had enough samples to train the DNN.  ~The users voice is turned into a series of waveform samples, spectrum analysis then converts the waveform sample stream into a sequence of frames. 20 of these frames (around 20 seconds of audio) are fed into an acoustic model, a deep neural network which converts the speech into a probability distribution over a set of speech sound classes. The DNN is built using standard back-propegation and the gradient descent algorithm to minimise the cost function over iterations.  Temporal integration is then used to determine a confidence score  a 1 would indicate the system is certain that Hey Siri was said while a 0 indicates the opposite certainty. Apple utilised neural network training toolkits like TensorFlow in the process.  ~Success is measured in terms of both False Activation Rate (FAR) and False Reject Rate (FRR). For FAR Apple measures when Siri was triggered unnecessarily and for FRR Apple keeps track of how many times Siri fails to activate when a user says Hey Siri. There is a trade off between the two as a more sensitive model means an increase in FAR and a decrease in FRR. Overall, Apple expects to see an FRR of about 8 per 100 hours and an FAR of 2 per 100 hours.  ~If this confidence score lies above a threshold then the assistant is activated. If on the other hand the confidence score lies between the lower threshold (indicating that hey siri was not said) and the upper threshold (indicating hey siri was said then the phone enters a more sensitive mode, such that Siri is much more likely to be invoked if the user utters the same phrase as before within a few seconds. Apple interprets this to mean that the first attempt by the user was unsuccessful and they wish to try again.  
135 ~201 ~Article in BCM titled "Predicting advanced coronary calcium using machine learning" http://blogs.biomedcentral.com/bmcseriesblog/2017/10/31/predicting-advanced-coronary-calcium-using-machine-learning/ ~Cihan Oguz ~Identify risk of developing coronary artery disease based on the genetic variation of a subject using single nucleotide polymorphisms as a signifier of genetic variation. ~Two separate data sets which were taken from previous studies from ClinSeq® and the Framingham Heart. These were composed of middle-aged Caucasian men due to their higher risk of advanced coronary artery disease. No indication the size of the sample sizes were given ~Random forests and neural networks were used to classify single nucleotide polymorphisms which are associated with coronary artery disease. ~21 of the 56 Single nucleotide polymorphisms were classified as being predictive indicators of coronary artery disease. Those Single nucleotide polymorphisms were then used to identify genes which they are associated with.  ~
64 ~143 ~http://nordic.businessinsider.com/denmarks-largest-bank-is-using-machine-learning-to-predict-the-customers-behavior--and-they-like-it-2017-11/ ~Danske Bank/Advanced Analytics Team ~To predict the needs and communications of its customers. ~Customer data of the bank ~Training the models to analyze the customer data. ~Advanced Analytics has achieved 62% better results than normal campaigns of Danske Bank. ~With this, in the future there won't be any unpaid mortgage etc. because Advanced Analytics will only give it to those whose can pay. 
100 ~126 ~https://health.ucsd.edu/news/releases/Pages/2017-10-25-machine-learning-detects-marketing-and-sale-of-opioids-on-twitter.aspx ~University of California San Diego researchers ~Produce technology that mined Twitter to identify entities illegally selling prescription opioids online. ~11 million tweets that had the keywords for 3 commonly prescired opioid drugs Percocet, OxyContin and Oxycodone ~A two-step process of obtaining themes, and filtering out unwanted tweets was carried out in three subsequent rounds of machine learning. ~2.3 million tweets contained relevant content.  ~Results suggest that this could be a viable methodology for use in big data substance abuse surveillance, data collection, and analysis in comparison to other studies that rely upon content analysis and human coding schemes. 
143 ~100 ~ http://www.techradar.com/news/internet/how-recommendation-algorithms-know-what-you-ll-like-1078924 ~ AMAZON ~  Present recomendations to the customers ~ Customer history: viewed, rated and purchased items. ~ Collaborative Filtering: Each customer is represented by a vector which holds postive or negative numbers. The number is positive if the customer bought or rated the item, the number is negative if the customer disliked or the number would be 0 if the customer has not interacted with the item. There were some variants applied which would add an overall popularity factor to some items. The algorithm would generate it's recommendations by obtaining a similarity value between the current customer and everyone else. This was done by calculating the angles between the vectors. ~ Simple unsophisticated recommendations could be generated for customers using Amazon. ~This would be very computationally intense and popularity factors could easily inflate the results if they weren't regularly update. 
103 ~81 ~http://news.mit.edu/2017/artificial-intelligence-aids-materials-fabrication-1106 ~MIT, the University of Massachusetts at Amherst, and the University of California at Berkeley ~To automate the development of processes for producing materials. Scientists and engineers could provide the name of a material and be provided with suggested recipes. ~Approximately 640,000 research papers. ~A combination of supervised and unsupervised techniques. Word2vec (an algorithm developed at Google which looks at the contexts in which words occur) was used to expand the data set. ~A ML system that can isolate the recipes for materials from a paper and categorise the words in those recipes (eg into pieces of equipment, quantities, conditions etc). ~No comment. 
246 ~137 ~WIRED article titled "AI-powered lip sync puts old words into Obama's new mouth" http://www.wired.co.uk/article/ai-lip-sync-barack-obama ~Ira Kemelmacher-Shlizerman, University of Washington ~Developed a machine learning algorithm that can turn audio clips into realistic, lip-synced videos. This can then be used to graft realistic mouth shapes onto the head of a person from another video. ~Videos of the former United States President Barack Obama was used to train the neural network. The reason for their choice of the former president was justified by the need for at least 14 hours of video to train the neural network.   ~For data processing, the article only refers to the training of a neural network to generate realistic mouth shapes. However, the article displays an image depicting the processing chain as including the use of a recurrent neural network. ~No accuracy assessment of the process was provided in the article. However, the video included in the article which demonstrated the results appeared quite convincing. ~Despite the results appearing quite convincing, this is still no proof of its success as those videos could have been cherry picked by the researchers to overstate their success. The title is also misleading as machine learning not ‘AI’ was employed in the research. The validation for using a former president as a data source being the volume of data is plausible, However, it is likely that he was chosen as it the idea of mimicking the leader of a powerful country would sell in the media. 
217 ~139 ~Cortex, IEEEXplore.  http://www.meetcortex.com/blog/how-machine-learning-can-save-the-whale/; http://ieeexplore.ieee.org/document/6968398/ ~Marineexplore and Cornell University.  Researchers: Soumya Sen Gupta and Sai Rajeshwar Indian Institute of Technology, Delhi India  ~The biggest threat to the Right Whale, which is an endangered species, is being hit by ocean vessels or getting caught in fishing nets.  The goal was to develop a system which can identify the call of the whales when they are nearby and so allow vessels to avoid them. ~The data used consisted of 3000 signals of 2 seconds duration from each of two classes: the Right Whale and other whales and creatures. ~The team then used wavelet packet decomposition techniques to find the best features to use to identify the whales calls.  To find the best features to use for classification they used sequential backward selection.  Following this, they used SVM and NaiveBayes classification algorithms to classify the signals.  A 10 fold crossvalidation accuracy test was carried out using both SVM and NaiveBayes to test the algorthims performances. ~The accuracies using SVM ranged from 78% to 90% and those of NaiveBayes ranged from 73% to 86% ~The team did not know a lot about the signals before beginning their work and felt that if they had more information about them beforehand they would have achieved higher accuracies  
182 ~494 ~http://scientific.cloud-journals.com/index.php/IJAFST/article/view/Sci-533 Safavian, S. R., & Landgrebe, D. (1991). A survey of decision tree classifier methodology. IEEE Trans. Systems, Man, & Cybernetics, 21(3), 660-674. Vidal, A., Talens, P., Prats-Montalban, J. M., Cubero, S., Albert, F., & Blasco, J. (2013). In-line estimation of the standard colour index of citrus fruits using a computer vision system developed for a mobile platform. Food and Bioprocess Technology, 6(12), 3412-3419. ~Nandan Thor, Master of Science student, California Polytechnic State University, San Luis Obispo ~Prediction ripeness of Bannannas to minimize food waste. ~46 bananas, were used. Every days each bananas was photgraphed. Exteral factors such as temp, humidity and distance to camera was kept constant day to day. RGB data was then extracted from the photos. ~K-means clustering, knn and other algorithim were used and then compared against one another. ~Results showed that bananas were at ripeness between days 26 and 31. Decision tree classifier was most effictive with 52% accuracy. Gaussian and knn followed with 51% and 50% accuracy respectively. ~Massive impication for making food industry more effiencent. It currently loses $15bn dollars to food waste every year. 
31 ~124 ~http://aclweb.org/anthology/P/P16/P16-1174.pdf?lipi=urn%3Ali%3Apage%3Ad_flagship3_pulse_read%3BOp4OMzswSRWHfZXZIJt33A%3D%3D ~Duolingo ~Predict thewords a student remembers and which they forget ~12.9 million data records from Duolingo users ~Prediction based on historical practice data, including frequency and difficulty ~Accurate prediction ~
35 ~72 ~https://www.digitaltrends.com/cool-tech/vicarious-ai-research-captchas/ ~Vicarious ~Create a program that will recognise CAPTCHA and successfully complete them ~CAPTCHA ~Uses a neural network and a hierarchal tree of related features to determine which letter is which ~Found the article interesting 
409 ~111 ~https://www.technologyreview.com/s/608777/why-googles-ai-can-write-beautiful-songs-but-still-cant-tell-a-joke/ ~MidiNet ~User study to compare the melody of eight-bar long generated by MidiNet and by Google’s MelodyRNN models, each time using the same priming melody ~In order to teach MidiNet before the comparisons were made, it crawled a collection of 1,022 MIDI tabs of pop music from TheoryTab, which provides exactly two channels per tab, one for melody and the other for the underlying chord progression. With this dataset, we can implement at least two versions of MidiNets, one that learns from only the melody channel for fair comparison with MelodyRNN, which does not use chords, and the other that additionally uses chords to condition melody generation,to test the capacity of MidiNet. Both MidiNet  and Google's MelodyRNN created a dataset used specific paramaters, core model, data type, genre specificity as well as mandatory knowledge that the machine had to know before generating any melodies, these were based on priming melody and the music scale and a melody profile. ~Symbolic Representation for Convolution, Generator CNN and Discriminator CNN, Conditioner CNN ~To evaluate the aesthetic quality of the generation result, a user study that involves human listeners was needed. They conducted a study with 21 participants. Ten of whom understand basic music theory and have the experience of being amateur musicians, so they were considered as people with musical backgrounds, the remaining 11 did not have a musical backgrounds at all. Midinet was compared with 3 models realesed by Google Magenta, 100 priming melodies were chosen and each had to create melodies of 8 bars following these primers. The users had to rate the music by 3 metrics, how pleasing, how real, how interesting on a scale of 1 to 5, MelodyRNN used the 3 models lookback RNN, attention RNN and basic RNN , both the lookback RNN and attention RNN outperformed the basic RNN in all 3 metrics and across both groups of users, With MidiNet model 1, this performed alot better than the 3 MelodyRNN models, especially