From 141aea0765d9559f85f311e989d7a1c5587e4fdf Mon Sep 17 00:00:00 2001 From: AbdulAyeshaSiddiqua <67275123+AbdulAyeshaSiddiqua@users.noreply.github.com> Date: Thu, 25 Jun 2020 23:15:34 +0530 Subject: [PATCH 1/3] headerlink3 --- src/lab/exp3/Objective.html | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/src/lab/exp3/Objective.html b/src/lab/exp3/Objective.html index df3ae615c..22bea7313 100644 --- a/src/lab/exp3/Objective.html +++ b/src/lab/exp3/Objective.html @@ -130,7 +130,7 @@

Perceptron learning

Community Links

- Sakshat Portal + Sakshat Portal

Outreach Portal @@ -227,4 +227,4 @@

Follow Us

- \ No newline at end of file + From a16832039eb57c87c6de82a9c0015d42669d53c7 Mon Sep 17 00:00:00 2001 From: AbdulAyeshaSiddiqua <67275123+AbdulAyeshaSiddiqua@users.noreply.github.com> Date: Mon, 29 Jun 2020 20:24:42 +0530 Subject: [PATCH 2/3] Delete :wq --- src/lab/exp1/:wq | 542 ----------------------------------------------- 1 file changed, 542 deletions(-) delete mode 100755 src/lab/exp1/:wq diff --git a/src/lab/exp1/:wq b/src/lab/exp1/:wq deleted file mode 100755 index 63eb61e46..000000000 --- a/src/lab/exp1/:wq +++ /dev/null @@ -1,542 +0,0 @@ - - - - - - - Artificial Neural Networks - - -
- - - -
- - - - - - - -
- - - -
- - - -
- - Parallel and distributed processing - I: Interactive activation and competition models - -
- - - - - -
- - - -
- -
- - -
- - -
- Objective -
- - -
-

-

The objective of IAC model is to illustrate the process of retrieving general and specific information from stored knowledge of specifics. -Some of the features of the human memory that are illustrated with this model are:

-
    -
      -
        -

      1. Retrieval by key (name) and by context.

        -

      2. Retrieval with noisy clues.

        -

      3. Assignment of default values and spontaneous generalization.

        -
      -
    -
-



-

- - - -
- - -
- - - - - - - -
- -
- - -
- -
- Tutorial -
- -
- -

PDP Model Development

-

-This model is illustrated by the help of Jets and Sharks database described in [McClelland, 1981] and also given in the following table. -

- - -

-The information about such a data, if stored in a computer memory, can be accessed by name or by any other set of items which can serve as a key to a particular set of information. -But this requires the method of access to be pre-programmed in the system. Moreover, certain characteristics of the data like the distribution of persons in different age groups or the -'common' characteristics among some persons, etc, can be obtained only by programming explicitly to derive the information embedded in the data. In other words, any information -in the data has to be sought explicitly. On the other hand human memory stores the information in the data in terms of patterns, and these patterns are useful to recall information even with partial clues. These features of the human memory can be demonstrated through representation in the form of a Parallel and Distributed Processing model, as shown in the figure below. -

- -

- - -
-
 Fig. 1: Figure illustrating the units and connections among those units in different pools. -Though here only some units are shown, but actually we have 27 units in the 'Name' pool. -Similarly there are 27 units in the 'Instance' pool, 2 in the 'Gang' pool and then 3 in rest of the other 4 pools.
- -

- -

-In the figure, the units are organized into different pools, such as 'Name' pool, 'Age' pool, etc. -The number of units in each pool corresponds to different possibilities in that category, -as for example, 27 units in the 'Name' pool and 3 units in the 'Age' pool, etc. -There are as many pools as there are categories (6 in this case), plus one additional pool called the 'Instance' pool. -

- -

PDP Model Description

- -

Units within each pool are connected in an inhibitory manner, i.e., the output of each unit is fed with a -negative weight to all other units in the same pool. The units in each pool -are connected to the corresponding units in the instance pool in an excitatory manner, i.e., the connection weights are positive. -For example, here, the 'Ralph' unit in the 'Name' pool, 'Jets' unit in the 'Gang' pool, 'in 30s' unit in the 'Age' pool, -'JH' unit in the 'Education' pool, 'pusher' unit in the 'Occupation' pool and 'Unmarried' unit in the 'Marital status' pool -are all connected to the 'Ralph' unit in the 'instance' pool with positive weights. -The units in the 'instance' pool are called "hidden" units, since by design, they are not accessible for any input or output. -The units in all other pools are "visible" units. Only 5 of 27 units are shown in the -'Name' pool and the 'instance' pool for illustration. Also, the inhibitory connections within each pool are not shown in the figure. -

- -

-There are a total of 68 units in the model. Each unit computes -a weighted sum of the input values to the unit fed from other units, -as well as from external inputs, if any. The weighted sum or the -activation value is passed through a nonlinear output function, which -gives as output the activation value itself, if the sum lies between -some prespecified minimum and maximum values. Otherwise, the -function gives either the minimum value at the lower limit or the -maximum value at the upper limit. The state of the model is described -as the collection of outputs of all the 68 units at any given instant of time. Starting -from any state, the next state can be computed by selecting a unit -at random and computing the weighted sum of its inputs first and -then the output of the unit. Due to change in the value of the output -of this unit, the model goes to a different state. Then another unit -is selected at random, and the new state for that unit is determined. -All the units are updated by selecting the units in a random sequence, -to compute one cycle of activation dynamics. After several cycles, the -model is guaranteed to reach a stable equilibrium state, when there -will not be any further change in the state of the model. -

- -

-For each set of external inputs, the model reaches a stable state -eventually. From the stable state the stored data can be read out. -For example, if we want the data about 'Ralph', the output (state) of the 'Ralph' unit in the 'Name' pool is set to maximum. Starting -with some initial values of state on other units, the network states are computed for -several cycles, until an equilibrium state is reached. At equilibrium, -there will be one unit in each pool having large positive value. Those -units correspond to the data that belongs to 'Ralph'. -

- -

Features Demonstrated By The Model

- -

-The model demonstrates several features of the functioning of the -biological neural network in human memory. Some of the features -are: (a) Retrieving an individual's data from his name, (b) retrieval -from a partial description, (c) graceful degradation, (d) default -assignment, and (e) spontaneous generalization for novel inputs. The -most important point is that the model stores the patterns embedded -in the given data. Therefore from the model, one could get even such -information for which the model was not explicitly designed. For -example, in the feature of default assignment, the model gives -possible good guess about missing information, even though we do -not know certain things about an individual. It evaluates the relative -strengths of the attributes from the given data in a complex manner, -which is difficult to describe explicitly. Thus, this model clearly brings -out the distinction between computer memory and human memory -for storage and retrieval of information. The model also brings out -the features of content-addressable memories and associative -memories for information retrieval. Note the distributed nature of the -memory in this model, in the sense that, the information is distributed -in the weights throughout the network. Also note the parallel and -distributed nature of the activation dynamics when the model is -realized in hardware, and also when it is allowed to relax naturally -changing from one state to another until an equilibrium state is -reached from the given initial state and external input. -Spontaneous generalization is the ability to provide a generalized picture of what is common to the memories that match the cues which are too vague to match a particular memory. -

- -
-
- - - -
- -
- -
- -
- Illustration -
- -
- -

Illustration of Interactive Activation and Competition Model

- -

-There are a total of 68 units in this model. Each unit computes a weighted sum of the input values to the unit, fed from other units as well as external inputs, if any. -The weighted sum or the activation value is passed through a non-linear output function, which gives as output, the activation value itself, if the sum lies between some -prespecified minimum and maximum values. Otherwise, the function gives either the minimum value at lower limit or the maximum value at the upper limit. The state of the model is described -as output of all the 68 units at any given instance of time. Starting from any state, the next state can be computed by selecting a unit at random and computing the weighted sum of its inputs first, and then the output of the function. -

-

-Due to change in the value of the output of this unit, the model goes to a different state. Then another unit is selected at random and the new state for that unit is determined. -All the units are updated by selecting all the units in a random sequence, to compute one cycle of activation dynamics. After several cycles, the model is guaranteed to reach -equillibrium state, when there will not be any further change in the state of the model. -

-

-For each set of external inputs, the model reaches a stable state eventually. From the stable state the stored data can be read out. -For example, if we want the data about 'Ralph', the output (state) of the 'Ralph' unit in the 'Name' pool is set to maximum. Starting -with some initial values of state on other units, the network states are computed for -several cycles, until an equilibrium state is reached. At equilibrium, there will be one unit in each pool having large positive value. -Those units correspond to the data that belongs to 'Ralph'. -

-

-In the experiment here, in the initial state we can see, as illustrated by the following figure, how 'Ralph' unit is connected in an inhibitory sense -(illustrated by red connections) with the rest of the members of the 'Name' pool. Also it is connected by an excitatory connection to the unit 'Ralph' in -the 'instance' pool (illustrated by a green connection). -

-

- -
-
- -

-

-Now, when an external input is applied by changing the state of the 'Ralph' unit (as by clicking the mouse over 'Ralph' node in the 'Name' pool) -we see that all the nodes settle down to a stable state after sometime. -

-

- -
-
-

-

-Finally when we try to read out the data related to 'Ralph' unit from the model, -we just need to visit every pool and find out which node has the highest positive value. -That data belongs to 'Ralph' unit. -

-

- -
-
-

-For this experiment, we can provide external bias and hence activation to units of all the pools except the instance pool. -The instance pool, which is also known as the hidden pool is not accessible to the end user. Yet we can see the connections -within and outside this pool by moving mouse pointer to respective units. -

-

-
- -
- - - - -
- -
- -
- -
- Procedure -
- -
-

-

    -
      -
        -

        -

      1. Check the initial state of the model by moving the mouse over various units in different pools.

        -

      2. Check the value for every unit and the inhibitory and excitatory connections it exhibits with units in same and different pools.

        -

      3. Click on any unit in the 'Name' pool and let the model settle down (in a couple of seconds). The settling down can be observed through the global change scale in the bottom margin.

        -

      4. Observe the connections between various units and pools (both green and red), popping up after the external input is provided.

        -

      5. Check with all the pools, the units with highest values within them, and match with the table provided in the tutorial section whether it belongs to same unit.

        -

        -
      -
    -
- -

-
- -
- - - - -
- -
- -
- -
- Experiment -
- -
-

- - - -
-

-
- -
- - - -
- -
- - -
- -
- Observations -
- -
- -

-

    -
      -
        -

      1. Observe the initial state of the model has inhibitory (red) connections with rest of the members in the same pool.

        -

      2. Note that the instance pool units have excitatory connections with all related data in every pool with that particular unit.

        -

      3. Note that initially all units in all the pools have negative activation value set at '-1.000' .

        -

      4. After clicking on any node in the 'Name' pool, we see that as the model settles down after calculation of weighted - inputs and outputs for respective units, all units have different values of output.

        -

      5. Observe the values for units in every pool and find out which unit has the highest value.

        -

      6. Click on units in other pools apart from 'Names' and observe how activating - multiple units leads to increased activation to units in the name pool.

        - -
      -
    -
- - -

- -
- -
- - - -
- -
- - -
- -
- Assignment -
- -
- -

-

-Using the experiment identify the following properties of human memory to store and retrieve knowledge: -

-
    -
      -
        -

      1. External activation to one unit in the name pool results in a positive change - in activation of units in other pools as well as the instance pool.

        -

      2. External activation to a unit in the name pool leads towards positive - activation of other names in the instance pool.

        -

      3. How the activation values changes when two or more names are chosen in the name pool.

        -
      -
    -
-

- -
- -
- - -
- -
- - -
-
- References -
-
-

-

    -
      -
        -

      1. B. Yegnanarayana, Artificial Neural Networks, New Delhi, India : Prentice-Hall of India, pg. 299, 1999.

        -

      2. J.L. McClelland, "Retrieving general and specific information from stored knowledge of specifics", in Proceedings of the Third - Annual Meeting of the Cognitive Science Society, pp. 170-172, 1981.

        -

      3. J.L. McClelland and D.E. Rumelhart, Explorations in Parallel Distributed Processing, Cambridge MA: MIT Press, 1988.

        -

      4. D.E. Rumelhart, G.E. Hinton, and R.J. Williams, "Learning internal representations by error propagation", in - Parallel Distributed Processing: Explorations in the Microstructure of Cognition, vol. 1 - (D.E. Rumelhart, J.L. McClelland, and the PDP Research Group, eds.), Cambridge, MA: MIT Press, pp. 318-362, 1986a.

        -
      -
    -
-

-
-
-
- - - - - - - -
- -
- -
- - - -
- -
- - - - -
- - - From b7ae0a620bc08635e9d3563b1027f84cd04d37a5 Mon Sep 17 00:00:00 2001 From: Ayesha Siddiqua Date: Mon, 29 Jun 2020 20:29:39 +0530 Subject: [PATCH 3/3] hide5 --- src/lab/exp1/Objective.html | 6 +++--- src/lib/js/MathJax | 1 - 2 files changed, 3 insertions(+), 4 deletions(-) delete mode 120000 src/lib/js/MathJax diff --git a/src/lab/exp1/Objective.html b/src/lab/exp1/Objective.html index 0f79a6b2d..138000dd7 100644 --- a/src/lab/exp1/Objective.html +++ b/src/lab/exp1/Objective.html @@ -127,13 +127,13 @@

Parallel and distributed processing - I: Interacti

Community Links

- Sakshat Portal +

- Outreach Portal +

- FAQ : Virtual Labs +

diff --git a/src/lib/js/MathJax b/src/lib/js/MathJax deleted file mode 120000 index 0a389d84f..000000000 --- a/src/lib/js/MathJax +++ /dev/null @@ -1 +0,0 @@ -mathjax \ No newline at end of file