GClasses::GNeuroPCA Class Reference

This class is a generalization of PCA. When the bias is clamped, and the activation function is "identity", it is strictly equivalent to PCA. By default, however, the bias is allowed to drift from the mean, which gives better results. Also, by default, the activation function is "logistic", which enables it to find non-linear components in the data. (GUnsupervisedBackProp is a multi-layer generalization of this algorithm.) More...

#include <GManifold.h>

Inheritance diagram for GClasses::GNeuroPCA:

Public Member Functions
	GNeuroPCA (size_t targetDims, GRand *pRand)
virtual	~GNeuroPCA ()
void	clampBias ()
	Specify to not update the bias values. More...
void	computeEigVals ()
	Specify to compute the eigenvalues during training. This method must be called before reduce is called. More...
double *	eigVals ()
	Returns the eigenvalues. Returns NULL if computeEigVals was not called. More...
virtual GMatrix *	reduce (const GMatrix &in)
	See the comment for GTransform::reduce. More...
void	setActivation (GActivationFunction *pActivation)
	Sets the activation function. (Takes ownership of pActivation.) More...
size_t	targetDims ()
	Returns the number of principal components to find. More...
GMatrix *	weights ()
	Returns the weight vectors. More...
Public Member Functions inherited from GClasses::GTransform
	GTransform ()
	GTransform (GDomNode *pNode, GLearnerLoader &ll)
virtual	~GTransform ()

Protected Member Functions
void	computeComponent (const GMatrix *pIn, GMatrix *pOut, size_t col, GMatrix *pPreprocess)
double	computeSumSquaredErr (const GMatrix *pIn, GMatrix *pOut, size_t cols)
Protected Member Functions inherited from GClasses::GTransform
virtual GDomNode *	baseDomNode (GDom *pDoc, const char *szClassName) const
	Child classes should use this in their implementation of serialize. More...

Protected Attributes
GActivationFunction *	m_pActivation
double *	m_pEigVals
GRand *	m_pRand
GMatrix *	m_pWeights
size_t	m_targetDims
bool	m_updateBias

Detailed Description

This class is a generalization of PCA. When the bias is clamped, and the activation function is "identity", it is strictly equivalent to PCA. By default, however, the bias is allowed to drift from the mean, which gives better results. Also, by default, the activation function is "logistic", which enables it to find non-linear components in the data. (GUnsupervisedBackProp is a multi-layer generalization of this algorithm.)

Constructor & Destructor Documentation

GClasses::GNeuroPCA::GNeuroPCA	(	size_t	targetDims,
		GRand *	pRand
	)

virtual GClasses::GNeuroPCA::~GNeuroPCA ( )

virtual

Member Function Documentation

void GClasses::GNeuroPCA::clampBias ( )

inline

Specify to not update the bias values.

void GClasses::GNeuroPCA::computeComponent ( const GMatrix *  pIn, GMatrix *  pOut, size_t  col, GMatrix *  pPreprocess  )

protected

void GClasses::GNeuroPCA::computeEigVals

(

)

Specify to compute the eigenvalues during training. This method must be called before reduce is called.

double GClasses::GNeuroPCA::computeSumSquaredErr ( const GMatrix *  pIn, GMatrix *  pOut, size_t  cols  )

protected

double* GClasses::GNeuroPCA::eigVals ( )

inline

Returns the eigenvalues. Returns NULL if computeEigVals was not called.

virtual GMatrix* GClasses::GNeuroPCA::reduce ( const GMatrix &  in )

virtual

See the comment for GTransform::reduce.

Implements GClasses::GTransform.

void GClasses::GNeuroPCA::setActivation

(

GActivationFunction *

pActivation

)

Sets the activation function. (Takes ownership of pActivation.)

size_t GClasses::GNeuroPCA::targetDims ( )

inline

Returns the number of principal components to find.

GMatrix* GClasses::GNeuroPCA::weights ( )

inline

Returns the weight vectors.

Member Data Documentation

GActivationFunction* GClasses::GNeuroPCA::m_pActivation

protected

double* GClasses::GNeuroPCA::m_pEigVals

protected

GRand* GClasses::GNeuroPCA::m_pRand

protected

GMatrix* GClasses::GNeuroPCA::m_pWeights

protected

size_t GClasses::GNeuroPCA::m_targetDims

protected

bool GClasses::GNeuroPCA::m_updateBias

protected