Matlab Neural Network

A simple and modular way of implementing a neural network with Matlab.

Layers

• FullyConnectedLayer(input_shape, output_shape)
• ConvolutionalLayer(input_shape, kernel_shape, layer_depth)
• MaxPoolLayer(input_shape, kernel_shape)
• FlattenLayer(input_shape)
• ActivationLayer(input_shape, activation)
• DropoutLayer(input_shape, drop_rate)

Activations

• Sigmoid ('sigmoid')
• Hyperbolic Tangant ('tanh')
• Rectified Linear Unit ('relu')
• Leaky Rectified Linear Unit ('leaky_relu')
• Linear ('linear')
• Exponential ('exp')
• Softplus ('softplus')
• Softsign ('softsign')

Losses

• Mean Squared Error ('mse')
• Mean Squared Logarithmic Error ('msle')
• Mean Absolute Error ('mae')
• Negative Logarithmic Likelihood ('neg_log_likelihood')
• Cross Entropy ('cross_entropy')

Example

Simple neural network applied to the XOR problem. See all examples here.

% IO data
input = [0 0 ; 0 1 ; 1 0 ; 1 1];
output = [0 ; 1 ; 1 ; 0];

% reshape data for neural network (sample dimension last)
input = reshape(rot90(input), [1,2, 4]);
output = reshape(rot90(output), [1,1, 4]);

% create a 3-layer neural network
net = Network({
    FullyConnectedLayer([1 2], [1 3])
    ActivationLayer([1 3], Activation("tanh"))
    FullyConnectedLayer([1 3], [1 1])
    ActivationLayer([1 1], Activation("tanh"))
});

% set cost function and learning_rate
net.build(Loss('mse'), 0.2)

% train on 1000 iterations
net.fit(input, output, 1000);

% test network
net.predict(input)

Save / Load Network

% save network
net.save('network.mat');

% load network
net = Network.load('network.mat');