NearNeural

Interface ThresholdingAlgorithm

All Known Implementing Classes:

HyperbolicTangentThresholdModel, LinearThresholdModel, SigmoidThresholdModel, SinusoidalThresholdModel, ThreePointRuleDifferentiation

public interface ThresholdingAlgorithm

A formal definition of an algorithm which adjusts input values.

Whilst a differential algorithm may not be used there is a very large chance of it happening - any capable backpropagation algorithm will attempt to minimise error by trying to locate the direction to move in which will reduce error gradient descent), and this is far easier if the differential is given more accurately, and this can be done best by hand. Extend ThreePointRuleDifferentiation if you wish to use numerical methods, although at the tiny descent levels usually used this is very inaccurate. If at all possible work out and code a differential algorithm manually, ThreePointRuleDifferentiation is very very bad by comparison.

Version:

0.3

Author:

Stephen Wattam

See Also:

SinusoidalThresholdModel, LinearThresholdModel, HyperbolicTangentThresholdModel, SigmoidThresholdModel

Method Summary

Methods

Modifier and Type	Method and Description
double	differentialValue(Edge[] edges) Returns the first differential of the value.
boolean	validateInput(double value) Returns true if the provided number is within reasonable input bounds for this algorithm.
double	value(Edge[] edges) Returns the value of a neuron.

Method Detail

value

double value(Edge[] edges)

Returns the value of a neuron. This is called by ReceptiveNeuron when it precalculates, and the value is stored in the neuron for later use, meaning that processing overheads can be offloaded onto a spare thread or into an otherwise idle part of the algorithm which manages the net/nets.

Parameters:

edges - All of the edges leading into this neuron

Returns:

The value of this neuron.

validateInput

boolean validateInput(double value)

Returns true if the provided number is within reasonable input bounds for this algorithm. This is to define input ranges which are acceptable for tha algorithm, but beware that edge values are randomly assigned from -1 to 1 during net initialisation: neurons may not take the exact value limits as typed, even on the first net. This last point is extremely important: inputs are very rarely fed in with a weight of 1.

Parameters:

value - The number to validate

Returns:

true if value is valid, false if not.

differentialValue

double differentialValue(Edge[] edges)

Returns the first differential of the value. This is used to work out the error gradient in a standard backkpropagation algorithm, although one may be implemented that omits this trait.

Parameters:

edges - The edges from which the value will be calculated

Returns:

The differential of that value