GClasses::GNeuralNetPseudoInverse Class Reference

Approximates the inverse of a neural network. (This only works well if the neural network is mostly invertible. For example, if the neural network only deviates a little from the identity function, then this will work well. With many interesting problems, this gives very poor results.) Note: This class assumes that the activation functions used within each layer of the neural network are homogeneous. More...

#include <GNeuralNet.h>

Public Member Functions
	GNeuralNetPseudoInverse (GNeuralNet *pNN, double padding=0.01)
	padding specifies a margin in which label values will be clipped inside the activation function output range to avoid extreme feature values (-inf, inf, etc.). More...
	~GNeuralNetPseudoInverse ()
void	computeFeatures (const double *pLabels, double *pFeatures)
	Computes the input features from the output labels. In cases of under-constraint, the feature vector with the minimum magnitude is chosen. In cases of over-constraint, the feature vector is chosen with a corresponding label vector that minimizes sum-squared error with the specified label vector. More...

Static Public Member Functions
static void	test ()

Protected Attributes
std::vector < GNeuralNetInverseLayer * >	m_layers
double	m_padding
double *	m_pBuf1
double *	m_pBuf2

Detailed Description

Approximates the inverse of a neural network. (This only works well if the neural network is mostly invertible. For example, if the neural network only deviates a little from the identity function, then this will work well. With many interesting problems, this gives very poor results.) Note: This class assumes that the activation functions used within each layer of the neural network are homogeneous.

Constructor & Destructor Documentation

GClasses::GNeuralNetPseudoInverse::GNeuralNetPseudoInverse	(	GNeuralNet *	pNN,
		double	padding = 0.01
	)

padding specifies a margin in which label values will be clipped inside the activation function output range to avoid extreme feature values (-inf, inf, etc.).

GClasses::GNeuralNetPseudoInverse::~GNeuralNetPseudoInverse

(

)

Member Function Documentation

void GClasses::GNeuralNetPseudoInverse::computeFeatures	(	const double *	pLabels,
		double *	pFeatures
	)

Computes the input features from the output labels. In cases of under-constraint, the feature vector with the minimum magnitude is chosen. In cases of over-constraint, the feature vector is chosen with a corresponding label vector that minimizes sum-squared error with the specified label vector.

static void GClasses::GNeuralNetPseudoInverse::test ( )

static

Member Data Documentation

std::vector<GNeuralNetInverseLayer*> GClasses::GNeuralNetPseudoInverse::m_layers

protected

double GClasses::GNeuralNetPseudoInverse::m_padding

protected

double* GClasses::GNeuralNetPseudoInverse::m_pBuf1

protected

double* GClasses::GNeuralNetPseudoInverse::m_pBuf2

protected