NearNeural

Class Edge

java.lang.Object

NearNeural.Edge

public class Edge
extends java.lang.Object

This represents an edge in a NearNeural net. It contains a reference to its source neuron and a weight. Thresholding algorithms may use these two in an arbitrary manner, but the usual method is to multiply an edge's source neuron value by the edge weight and sum all edges contained within the neuron, then pass it through a soft limiter function such as a sigmoid function.

Edges point from a downstream neuron upstream. This has been done to optimise calculation of values, but results in a slightly more ineficient backpropagation method where gradient descent is involved. Other backpropagation methods may not be hit severely by this, but it is an important consideration when designing the algorithm.

The edge's old weight is used when backpropagating

Version:

0.1.0a

Author:

Stephen Wattam

See Also:

Neuron

Field Summary

Fields

Modifier and Type	Field and Description
protected Neuron	source Stores a reference to the source neuron.

Constructor Summary

Constructors

Modifier	Constructor and Description
protected	Edge(Neuron p_source, double p_weight) Creates a new edge with neuron and weight.

Method Summary

Methods

Modifier and Type	Method and Description
double	getOldWeight() Returns the old weight, as per backpropagation.
Neuron	getSource() Retrns the target of this edge, ie the source (the neuron from which data flows unto the one which owns this Edge), but not the source of data.
double	getWeight() Returns the weight of the Edge.
void	setOldWeight(double p_weight) Set the old weight of this Edge.
protected void	setWeight(double p_weight) Sets the weight of this edge

Methods inherited from class java.lang.Object

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

Field Detail

source

protected Neuron source

Stores a reference to the source neuron. This is the neuron pointed at by the edge from a neuron downtsream from it. Edges may appear to point backwards, but this is in fact the manner in which they are most often used. Storing only a single-way link prevents much costly reference management code.

See Also:

setSource(Neuron)

Constructor Detail

Edge

protected Edge(Neuron p_source,
    double p_weight)

Creates a new edge with neuron and weight. The neuron given is the source of data, which has a target of whatever neuron to which this new edge is added.

Parameters:

p_source - The source neuron

p_weight - The weight of this edge

See Also:

Neuron

Method Detail

setOldWeight

public void setOldWeight(double p_weight)

Set the old weight of this Edge. This is used during backpropagation to simulate a double-pass method, as backpropagation calculations ought to use this to simulate a lack of change in a downstream neuron.

Parameters:

p_weight - The weight to set

See Also:

BackPropagationMethod

getOldWeight

public double getOldWeight()

Returns the old weight, as per backpropagation.

Returns:

The old weight of this edge

getSource

public Neuron getSource()

Retrns the target of this edge, ie the source (the neuron from which data flows unto the one which owns this Edge), but not the source of data.

Returns:

the Neuron this edge moves data from.

setWeight

protected void setWeight(double p_weight)

Sets the weight of this edge

Parameters:

p_weight - The weight to set this edge to

getWeight

public double getWeight()

Returns the weight of the Edge.

Returns:

The weight of this Edge