Procedures

Given input x and output y, evaluates the position of the maximum value of the output and returns the number of matches relative to the size of the dataset.

Overloads the default type constructor.

Overloads the default type constructor.

Applies a backward propagation through the network and returns the weight and bias gradients.

Layer class constructor. this_size is the number of neurons in the layer. next_size is the number of neurons in the next layer, used to allocate the weights.

Performs a collective sum of bias tendencies.

Initialises biases structure.

Returns an array of 10 reals, with zeros everywhere and a one corresponding to the input number, for example: digits(0) = [1., 0., 0., 0., 0., 0., 0., 0., 0., 0.] digits(1) = [0., 1., 0., 0., 0., 0., 0., 0., 0., 0.] digits(6) = [0., 0., 0., 0., 0., 0., 1., 0., 0., 0.]

Performs a collective sum of weights tendencies.

Initialises weights structure.

Performs the forward propagation and stores arguments to activation functions and activations themselves for use in backprop.

Gaussian activation function.

First derivative of the Gaussian activation function.

Allocates and initializes the layers with given dimensions dims.

Converts an array of MNIST labels into a form that can be input to the network_type instance.

Loads the network from file.

Loads the MNIST dataset into arrays.

Given input x and expected output y, returns the loss of the network.

Network class constructor. Size of input array dims indicates the total number of layers (input + hidden + output), and the value of its elements corresponds the size of each layer.

Use forward propagation to compute the output of the network. This specific procedure is for a batch of 1-d input data.

Use forward propagation to compute the output of the network. This specific procedure is for a single sample of 1-d input data.

Prints a single image and label to screen.

Generates n random numbers with a normal distribution.

Generates m x n random numbers with a normal distribution.

REctified Linear Unit (RELU) activation function.

First derivative of the REctified Linear Unit (RELU) activation function.

Saves the network to a file.

Sets the activation function. Input string must match one of provided activation functions, otherwise it defaults to sigmoid. If activation not present, defaults to sigmoid.

A thin wrapper around layer % set_activation(). This method can be used to set an activation function for all layers at once.

A thin wrapper around layer % set_activation(). This method can be used to set different activation functions for each layer separately.

Sigmoid activation function.

First derivative of the sigmoid activation function.

Step activation function.

First derivative of the step activation function.

Broadcasts network weights and biases from specified image to all others.

First derivative of the tanh activation function.

Tangent hyperbolic activation function. Same as the intrinsic tanh, but must be defined here so that we can use procedure pointer with it.

Given input global array size, return start and end index of a parallel 1-d tile that correspond to this image. start and end indices assuming equal tile sizes if we have any remainder, distribute it to the tiles at the end

Trains a network using input data x and output data y, and learning rate eta. The learning rate is normalized with the size of the data batch. mini-batch size number of layers

Trains for num_epochs epochs with mini-bachtes of size equal to batch_size.

Trains a network using a single set of input data x and output data y, and learning rate eta.

Updates network weights and biases with gradients dw and db, scaled by learning rate eta. update biases update weights