--------------------------------------------------------------------------
__
| \|\ /_ _ _ _ _ |_
|__/| \/(-| (_)(-| )|_ Written by Nolan Conaway
_/ bingweb.binghamton.edu/~nconawa1/
/\ |_ _ _ _ _ _ _| _ _ Updated January 4, 2016
/--\|_||_(_)(-| )(_(_)(_|(-|
--------------------------------------------------------------------------
This set of scripts runs a minimal version of the DIVA model of category learning (Kurtz, 2007). It is written in MATLAB, and is currently set up to simulate the six-types problem of Shepard, Hovland, & Jenkins (1961)--though it should generalize to any dataset. There are a variety of utility scripts, and a few important ones:
- START.m can be used to test DIVA using particular parameter sets.
- DIVA.m uses a provided architecture to train a network on a set of inputs and category assignments.
- FORWARDPASS.m and BACKPROP.m are used to propagate activations forward through the model, and error backward through the model, respectively. BACKPROP.m additionally completes a weight update.
- responserule.m implements DIVA's response rule, including its new focusing component.
Simulations are run by executing the START.m script. All simulations begin by passing a model struct to the DIVA.m function. At a minimum, 'model' needs to include:
| Field | Description | Type |
|---|---|---|
inputs |
Training items | Item-by-feature matrix |
labels |
Class assignments | Column integer vector |
numblocks |
# of passes through the training set | Integer (>0) |
numinitials |
# of random initial networks | Integer (>0) |
outputrule |
activation rule for the category channels | 'sigmoid' or 'linear' |
numhiddenunits |
number of hidden units | Integer (>0) |
learningrate |
Learning rate | Float (0 - Inf) |
weightrange |
range of initial weight values | Float (0 - Inf) |
betavalue |
Focusing Parameter | Float (0 - Inf) |
For almost all situations, inputs should be scaled to [-1 +1]. However, the target activations should be scaled to [0 1], in order to permit logistic output units. By default, the program automatically computes targets as scaled versions of the inputs (targets = inputs / 2 + 0.5). This is done in DIVA.m
By default, DIVA uses linear output units for continuous datasets, and logistic outputs for binary datasets. This option is set using outputrule: 'sigmoid' for logistic units, other values will result in linear. When linear outputs are used, SSE will be the error measure. Cross-entropy error is used for logistic outputs. Hidden units are always logistic.
DIVA.m will train the network and return a result struct. As-is, 'result' contains only training accuracy for each initialization at each training block. Additional measures, such as test phase classification, can be added. You will need to write custom code to compare DIVA's performance to a set of behavioral data.