Machine-Learning-Projects

A collection of Machine Learning projects I completed for coursework in my 4th year at university.

CNN - Image Classification with the Mars Curiosity Rover

This project focused on classifying images taken by the Curiosity Mars Rover into one of 25 distinct categaroies using a Convolutional Neural Network (CNN). After this, the model was changed to perform binary classification and categorise the images based on if they contained a part of the rover or not.

Data	Link
Mars surface image (Curiosity rover) labeled dataset	NASA's Open Data Portal

• Utilised ImageDataGenerator to load and rescale the image data, and prepared batches for training while also exploring and pre-processing the dataset.
• Implemented a CNN model (using TensorFlow and Keras libraries to perform the classification task) to train on 3746 images and experimented with the network architecture for this particular problem.
• Optimised hyperparameters such as batch size and dropout to improve accuracy while reducing runtime.
• Evaluated the initial model's performance with accuracy and loss plots, and a histogram showing the distribution of Correct vs Incorrect predictions per class.

CNN - Image Classification from a Kaggle-sourced anime dataset

This project also classified images; however, these images were from a subset of popular anime titles in a Kaggle dataset. For training, validation and testing, 6573, 1618 and 2040 images were used, respectively; this was a collection of approximately 300 images representing each of the 28 chosen classes.

Data	Link
Original dataset	Kaggle
Modified dataset (Used in this project)	Google Drive

• Developed and trained a CNN to perform classification, labelling each of the training images with a predicted class from the 28 available.
• Combined part of the image pre-processing into the CNN with a lamda layer in between the input and first convolutional layers, resizing each image as it is put into the network.
• Experimented with the network architecture and hyperparameters to result in a CNN optimised for this dataset and task.
• Evaluated the model's performance through a combination of:
  • Accuracy and Loss plots
  • Histograms of the Correct vs. Incorrect label predictions per class
  • Line graphs of the model's percentage accuracy for each class
  • Confusion matrices

MLP - Prediction of z-band brightness of quasars, given the brightness in i- and r-bands

This project used machine learning to predict a quasar's brightness in the z-band after training a Multi-Layer Perceptron (MLP) on the quasar's brightness in the i- and r-bands.

Data	Link
PennState Center for Astrostatistics	Data & Tutorials
SDSS quasar catalog	SDSS_quasar.dat

• Created a heatmap of the dataset to visualise correlations between different variables, choosing the optimal set to perform multi-variable regression with.
• Developed and trained an MLP on the selected data after pre-processing, allowing the network to take in the i- and r-band magnitudes.
• Tuned hyperparameters like batch size, learning rate, dropout and hidden layers to optimise the network performance.
• Evaluated the model using Mean Squared Error (MSE), achieving an MSE loss of 0.072, and residual analysis on the test data.