rasumovsky/learningproject

A feed-forward neural net with back-propagation learning.

LearningProject

This package is just a playground for learning about neural networks. The idea is to make a feed-forward back-propagational neural network. The test case will be separation of gluon-fusion produced Higgs to gamma gamma signals from non-resonant diphoton backgrounds.

Axon

A basic class for representing a directed weighted connection between two nodes.

Neuron

A basic class for representing network nodes. This includes the activation function, the incoming connections, and the outgoing connections. There is also the option to create the node as a bias node, with a constant response of 1.0.

NeuralNetwork

A class for creating a network with a given number of variables, a given number of hidden layers, and a given number of nodes per hidden layer. The first (visible) layer should have a linear activation function, and the input variables should be transformed to the interval [-1,+1], preferrably in a uniform manner (whatever that means). The output layer(s) should also have linear activation functions, so that the output is on the interval [-1,+1].

The response of the network can be found simply by calling the getResponse() method of the output layer node(s). This recursively calls the same method for upstream neurons, if the output has not already been calculated following the most recent call to the clearResponse() method.