nlgranger / nnbox

A neural network toolbox for matlab

NNBox

NNBox is a Matlab © toolbox for neural networks. Many other toolboxes are already available for matlab and may either offer more models, a higher levels of support, better optimization, or simply a bigger user community... This toolbox tries to focus on different objectives:

• Providing very clear and simple implementations of some neural networks models and architectures.
• Providing a simple template to implement new models rapidly
• Providing a flexible interface where building blocks can be arranged together easily.

In particular, this library provides support for Restricted Boltzmann Machines (RBM), Convolutional Neural Networks (CNN), simple perceptrons models. It allows to arrange these models in parallel, as stacked multiple layers, or even in a Siamese network architecture.

This library does not focus on completeness though, because attempting to do so rarely gives satisfying results. Instead it tries to provide simple and flexible architectural fundations to help you implement your own model quickly.

For your information, here is a list of other existing libraries:

• Matlab Neural Network toolbox
• DeepLearnToolbox A popular deep learning toolbox
• MEDAL Similarily provides implementations for several sorts of Deep Learning models.
• MatConvNet Provides awrapper to a C++ implementation of convolutional neural networks. It is actually used here for the CNN model.

Requirements

As far as I can tell, any version of matlab above R2011a should work, R2014a is known to work. Octave is not supported because classes are not yet fully supported.

Installation

Just add nnbox subfolders to your path:

addpath('nnbox/utils:nnbox/networks:nnbox/costfun:nnbox/distances');

CNN implementation requires the MatConvNet library as a backend, follow installation instructions and add the matlab bindings to the path.

Examples

X = [0  1  0  1;
     0  0  1  1];
Y = [0 .5 .5  1];
net = Perceptron(2, 1, struct('lRate', 0.5));
trainOpts = struct('nIter', 100, 'displayEvery', 10);
train(net, SquareCost(), X, Y, trainOpts);

• MNIST figure recognition using a Deep belief network : examples/MNIST_DNN.m

Documentation

Refer to DOCUMENTATION.md