jsat.classifiers.neuralnetwork.regularizers

Class Max2NormRegularizer

java.lang.Object

jsat.classifiers.neuralnetwork.regularizers.Max2NormRegularizer

All Implemented Interfaces:

Serializable, WeightRegularizer

public class Max2NormRegularizer
extends Object
implements WeightRegularizer

This regularizer restricts the norm of each neuron's weights to be bounded by a fixed constant, and rescaled when the norm is exceeded.

Author:

Edward Raff

See Also:

Serialized Form

Constructor Summary

Constructors

Constructor and Description
Max2NormRegularizer(double maxNorm)

Method Summary

Modifier and Type	Method and Description
void	applyRegularization(Matrix W, Vec b) Applies regularization to one matrix, where the rows of the matrix correspond tot he weights associated to one neuron's input.
void	applyRegularization(Matrix W, Vec b, ExecutorService ex) Applies regularization to one matrix, where the rows of the matrix correspond tot he weights associated to one neuron's input.
double	applyRegularizationToRow(Vec w, double b) Applies the regularization to one row of the weight matrix, where the row corresponds to the weights into one neuron.
Max2NormRegularizer	clone()
double	getMaxNorm()
void	setMaxNorm(double maxNorm) Sets the maximum allowed 2 norm for a single neuron's weights

Methods inherited from class java.lang.Object

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

Constructor Detail

Max2NormRegularizer

public Max2NormRegularizer(double maxNorm)

Method Detail

setMaxNorm

public void setMaxNorm(double maxNorm)

Sets the maximum allowed 2 norm for a single neuron's weights

Parameters:

maxNorm - the maximum norm per neuron's weights

getMaxNorm

public double getMaxNorm()

Returns:

the maximum allowed 2 norm for a single neuron's weights

applyRegularization

public void applyRegularization(Matrix W,
                                Vec b)

Description copied from interface: WeightRegularizer

Applies regularization to one matrix, where the rows of the matrix correspond tot he weights associated to one neuron's input. The vector of bias terms must then have the same length as the number of rows in the given matrix.

Specified by:

applyRegularization in interface WeightRegularizer

Parameters:

W - the matrix to apply regularization to

b - the vector of bias terms to apply regularization to

applyRegularization

public void applyRegularization(Matrix W,
                                Vec b,
                                ExecutorService ex)

Description copied from interface: WeightRegularizer

Applies regularization to one matrix, where the rows of the matrix correspond tot he weights associated to one neuron's input. The vector of bias terms must then have the same length as the number of rows in the given matrix.

Specified by:

applyRegularization in interface WeightRegularizer

Parameters:

W - the matrix to apply regularization to

b - the vector of bias terms to apply regularization to

ex - the source of threads for parallel computation

applyRegularizationToRow

public double applyRegularizationToRow(Vec w,
                                       double b)

Description copied from interface: WeightRegularizer

Applies the regularization to one row of the weight matrix, where the row corresponds to the weights into one neuron.

Specified by:

applyRegularizationToRow in interface WeightRegularizer

Parameters:

w - the weight row to be altered depending on the regularization method

b - the original bias input to this row

Returns:

the new bias value, or the same value if no change in the bias has occurred

clone

public Max2NormRegularizer clone()

Specified by:

clone in interface WeightRegularizer

Overrides:

clone in class Object