This is an implementantion in Python of a simple model of a neural network. It uses the sigmoid function as the activation function, stochastic gradient descent for learning, and quadratic cost function (cross-entropy should be available soon). Weights and bias are initialized from a gaussian distribution (when not loaded), and are choosen using a low standart deviation to prevent neurons from being initially saturated.

You can initialize the network with random weights and biases, and i input neurons, h hidden layers and o output layers like so (any number of hidden layers is allowed):

NN = NeuralNetwork([i, h, o])

You can also initialize the net with a set of previously saved weights and bias by passing a preset name to the constructor (which should be present in the presets folder):

NN = NeuralNetwork("my_cool_preset")

Learning can be achieved calling the stochastic_gradient_descent method:

NN.stochastic_gradient_descent(training_data, test_data, epochs=10, batch_size=10, learning_rate=3.0)

training_data (and test_data, if needed) should be a list of of tuples (x, y). x should be a numpy column vector with i elements, i being the number of neurons in the input layers, and y should also be a column vector with o elements, where o is the number of neurons in the output layer.

If test_data is provided, at each epoch the method will print the number of itens in test_data that the network got correct.

One can compute the output of the network as easily as

NN.feed_forward(x)

x is the input, and should be a numpy column vector with the same number of entries and the number of input neurons in the input layer.

It is possible to load the MNIST data using the following methods:

handwritten_digits = MNIST("data/mnist")
training_data = handwritten_digits.load_training()
testing_data = handwritten_digits.load_testing()

You can also save the current set of weights and biases in a external file, in the presets folder:

NN.save_parameters("my_cool_preset")