Model properties file format

This document specifies the format used for defining combinations of model properties.

Introduction

This format uses the JSON notation, with extra support for comments notated with // ... and /* ... */.

An example of file that follows this format is available in config/model/final.json.

Format

The root element of the file must be an array ([...]) that contains one object ({...}) for each model properties combination that is desired to test.

The file has an structure like this:

[
    {...},
    {...},
    ...
]

A model has the following properties:

• model ("hopfield" or "restricted-boltzmann") indicates the type of model to test.
• If model is "hopfield", these keys must be set:
  • trainingRule ("hebbian" or "storkey") defines the training rule to use.
• If model is "restricted-boltzmann", these keys must be set:
  • seed (integer) defines a seed for the random generator, to reproduce experiments.
  • hiddenNeurons (integer >= 0) indicates the number of hidden neurons to use.
  • learningRate (float > 0) indicates the learning rate for the training.
  • weightDecay (float >= 0) indicates the weight decay for the training.
  • momentum (boolean) indicates if momentum is used during the training.
  • batchSize (integer > 0) indicates the batch size for the training. Use 1 for online training.