README.md

Neural Networks in MATLAB

Experiments with Neural Networks in MATLAB.

a = input;

for j=1:numel(layers)  
    weights =  layers{j}.theta;
    sigma   =  layers{j}.sigma;
    
    z = weights * [1; a];
    a = sigma(z);
end

output = a;

Start with ann.m. This code implements a training example and utilizes the following functions:

• feedforward.m: Runs inputs through the neural network, producing the hypothesis of the result.
• feedforward_for_training.m: Like feedforward.m, but produces additional output required for the backpropagation stage.
• backpropagate.m: Performs the error backpropagation and produces weight deltas.

Depending on the configuration, one of the following Gradient Descent algorithms can be used:

• momentum_gradient_descent.m: This is regular gradient descent with a fixed learning rate that additionally uses a momentum term to add in fractions of the previous gradient.
• accelerated_gradient_descent.m: This implementation utilizes adaptive (i.e. different) delta values per weight parameter instead of a single learning rate and only uses the gradient to check for direction changes.

In all cases, the concept of flat spot elimination (FSE) can be used to aid in navigating areas of small error gradients by artificially providing an empirically set gradient direction to move along. Small values such as 0.1 might work, disable by setting it to 0.