org.cognity.neural.train

Class Backpropagation

java.lang.Object

org.cognity.neural.train.GradientDescent

org.cognity.neural.train.Backpropagation

All Implemented Interfaces:

IterativeProcess, TrainingRule

public class Backpropagation
extends GradientDescent

The backpropagation algorithm with momentum. This algorithm works only with multilayer perceptrons. Furthermore, the activation function in each layer is required to be differentiable. This algorithm requires a supervised data set.

The weights are updated according to the following formula:

Where:

• is the weight matrix of the link
• is the learning rate
• is the delta vector of the target layer of the link
• is the input vector of the target layer of the link
• is the momentum constant
• is the last update of the weight matrix

The weights are updated after every pattern presentation (no batch mode available yet). The measured error is the total error for one epoch, that is:

where P is the number of data examples.

Author:

Artur Markiewicz

Field Summary

Fields inherited from class org.cognity.neural.train.GradientDescent

DEFAULT_LEARNING_RATE, DEFAULT_MOMENTUM

Constructor Summary

Constructors

Constructor and Description
Backpropagation(FeedforwardNet network) Creates a new object using default learning rate and momentum values.
Backpropagation(FeedforwardNet network, double learningRate, double momentum) Creates a new object using provided learning rate and momentum values.

Method Summary

Methods

Modifier and Type	Method and Description
FeedforwardNet	getNetwork() Returns a network trained by this rule.
protected void	trainExample(DataExample example) Learns a single pattern.

Methods inherited from class org.cognity.neural.train.GradientDescent

getCurrentError, getCurrentIteration, getLastError, getLearningRate, getMomentum, getStopCondition, setCurrentError, setLearningRate, setMomentum, setStopCondition, train

Methods inherited from class java.lang.Object

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

Constructor Detail

Backpropagation

public Backpropagation(FeedforwardNet network)

Creates a new object using default learning rate and momentum values.

Backpropagation

public Backpropagation(FeedforwardNet network,
               double learningRate,
               double momentum)

Creates a new object using provided learning rate and momentum values.

Parameters:

learningRate - the learning rate value

momentum - the momentum value

Method Detail

getNetwork

public final FeedforwardNet getNetwork()

Description copied from interface: TrainingRule

Returns a network trained by this rule.

Returns:

the network.

trainExample

protected final void trainExample(DataExample example)

Learns a single pattern.

Specified by:

trainExample in class GradientDescent

Parameters:

example -