de.jungblut.classification.nn

Class RBMCostFunction

java.lang.Object

de.jungblut.math.minimize.AbstractMiniBatchCostFunction

de.jungblut.classification.nn.RBMCostFunction

All Implemented Interfaces:

CostFunction

public final class RBMCostFunction
extends AbstractMiniBatchCostFunction

Restricted Boltzmann machine implementation using Contrastive Divergence 1 (CD1). This algorithm is based on what has been teached by Prof. Hinton in the Coursera course "Neural Networks for Machine Learning" '12. This is an unsupervised learning algorithm to train high level feature detectors. NOTE: Sutskever and Tieleman have shown that it is not following the gradient of any function (Sutskever and Tieleman, 2010). So this isn't minimizable using line searching optimizers like Fmincg, GradientDescent is doing a great job though as it doesn't care if you're moving into the right direction down hill.

Author:

thomas.jungblut

Constructor Summary

Constructors

Constructor and Description
RBMCostFunction(de.jungblut.math.DoubleVector[] currentTrainingSet, int batchSize, int numThreads, int numHiddenUnits, ActivationFunction activationFunction, TrainingType type, double lambda, long seed, boolean stochastic)

Method Summary

Modifier and Type	Method and Description
protected CostGradientTuple	evaluateBatch(de.jungblut.math.DoubleVector input, de.jungblut.math.DoubleMatrix data, de.jungblut.math.DoubleMatrix outcomeBatch) Evaluate the batch.

Methods inherited from class de.jungblut.math.minimize.AbstractMiniBatchCostFunction

evaluateCost

Methods inherited from class java.lang.Object

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

Constructor Detail

RBMCostFunction

public RBMCostFunction(de.jungblut.math.DoubleVector[] currentTrainingSet,
                       int batchSize,
                       int numThreads,
                       int numHiddenUnits,
                       ActivationFunction activationFunction,
                       TrainingType type,
                       double lambda,
                       long seed,
                       boolean stochastic)

Method Detail

evaluateBatch

protected CostGradientTuple evaluateBatch(de.jungblut.math.DoubleVector input,
                                          de.jungblut.math.DoubleMatrix data,
                                          de.jungblut.math.DoubleMatrix outcomeBatch)

Description copied from class: AbstractMiniBatchCostFunction

Evaluate the batch.

Specified by:

evaluateBatch in class AbstractMiniBatchCostFunction

Parameters:

input - the parameters to use.

data - the batch matrix as input (already contains a bias!).

outcomeBatch - the batch matrix denoting the output.

Returns:

the cost/gradient tuple usually used when using AbstractMiniBatchCostFunction.evaluateCost(DoubleVector).