NearNeural

Class ThreadedUnweightedBackPropagationMethod

java.lang.Object

NearNeural.UnweightedBackPropagationMethod

NearNeural.ThreadedUnweightedBackPropagationMethod

All Implemented Interfaces:

BackPropagationMethod

public class ThreadedUnweightedBackPropagationMethod
extends UnweightedBackPropagationMethod
implements BackPropagationMethod

A threaded version of UnweightedBackPropagationMethod which uses a standard gradient-descent backpropagation method. One thread per passed neuron is started, each of which recurses and follows all paths through the net. This results in some large overheads for thread management, but allows for faster training if using a cluster or multi-core machine. It is 100% algorithmically compatible with UnweightedBackPropagationMethod and can be swapped out without re-training destroying the net's abilities.

The only instance in which this is worth using instead of an unthreaded backpropagation method is where many threads may run in parallel (such as a quad-core chip) and where each output neuron is likely to have many edges leading into it. The overheads involved in the creation of a thread per neuron are worthwhile only when many edges need to be processed per neuron.

Version:

0.3

Author:

Stephen Wattam

Field Summary

Fields inherited from class NearNeural.UnweightedBackPropagationMethod

learningRate

Constructor Summary

Constructors

Constructor and Description
ThreadedUnweightedBackPropagationMethod(double p_learningRate) Creates a new ThreadedUnweightedBackPropagationMethod with the given learning rate.

Method Summary

Methods

Modifier and Type	Method and Description
void	backPropagateHidden(NeuronLayer downstream, ReceptiveNeuron[] neurons) Back propagates a hidden layer relative to the layer immediately following it.
double	backPropagateOutputs(double[] desirable, ReceptiveNeuron[] neurons) This function back-propogates output nodes.

Methods inherited from class NearNeural.UnweightedBackPropagationMethod

getLearningRate, setLearningRate

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

ThreadedUnweightedBackPropagationMethod

public ThreadedUnweightedBackPropagationMethod(double p_learningRate)

Creates a new ThreadedUnweightedBackPropagationMethod with the given learning rate.

Parameters:

p_learningRate - The learning rate to use.

Method Detail

backPropagateHidden

public void backPropagateHidden(NeuronLayer downstream,
                       ReceptiveNeuron[] neurons)

Back propagates a hidden layer relative to the layer immediately following it.

Specified by:

backPropagateHidden in interface BackPropagationMethod

Overrides:

backPropagateHidden in class UnweightedBackPropagationMethod

Parameters:

downstream - The layer downstream from this

neurons - The neurons in the layer to backprop.

backPropagateOutputs

public double backPropagateOutputs(double[] desirable,
                          ReceptiveNeuron[] neurons)

This function back-propogates output nodes. It is subtly different to other Neuron layers' backprop algorithm and is called first by NeuralNet. It returns sum([delta]o/n), the sum of all error functions with respect to weightings on each output node. This returned value must be passed to the other node layers during the second stage of backprop training.

Specified by:

backPropagateOutputs in interface BackPropagationMethod

Overrides:

backPropagateOutputs in class UnweightedBackPropagationMethod

Parameters:

desirable - The desirable set of inputs to adapt to

neurons - The neurons to backpropagate

Returns:

The mean squared error for the whole net.

Throws:

IndexOutOfBoundsException - in the event that the number of desirable outputs does not match that of the net.