fr.lip6.jkernelmachines.kernel.extra

Class NystromKernel<T>

java.lang.Object

fr.lip6.jkernelmachines.kernel.Kernel<T>

fr.lip6.jkernelmachines.kernel.extra.NystromKernel<T>

All Implemented Interfaces:

java.io.Serializable

public class NystromKernel<T>
extends Kernel<T>

This kernel provides a fast approximation of a given kernel using the Nystrom approximation.

A fast active learning algorithm analog to that used in:
Fast Approximation of Distance Between Elastic Curves using Kernels
Hedi Tabia; David Picard; Hamid Laga; Philippe-Henri Gosselin
British Machine Vision Conference, Sep 2013, United Kingdom. British Machine Vision Conference

Author:

picard

See Also:

Serialized Form

Field Summary

Fields inherited from class fr.lip6.jkernelmachines.kernel.Kernel

name

Constructor Summary

Constructors

Constructor and Description
NystromKernel(Kernel<T> kernel) Default constructor with kernel to approximate as argument

Method Summary

Modifier and Type	Method and Description
void	activeTrain(java.util.List<TrainingSample<T>> list, int iterations, int samples, int pool)
java.util.List<TrainingSample<double[]>>	projectList(java.util.List<TrainingSample<T>> l)
double[]	projectSample(T sample) Project a sample to the space induced by the Nystrom approx
void	train(java.util.List<TrainingSample<T>> list) Train the Nystrom approx on a full training set.
double	valueOf(T t1) kernel similarity to zero
double	valueOf(T t1, T t2) compute the kernel similarity between two element of input space

Methods inherited from class fr.lip6.jkernelmachines.kernel.Kernel

getKernelMatrix, getKernelMatrixLine, getNormalizedKernelMatrix, normalizedValueOf, setName, toString

Methods inherited from class java.lang.Object

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

Constructor Detail

NystromKernel

public NystromKernel(Kernel<T> kernel)

Default constructor with kernel to approximate as argument

Parameters:

kernel - the kernel to be approximated

Method Detail

train

public void train(java.util.List<TrainingSample<T>> list)

Train the Nystrom approx on a full training set.

Might be costly for large training sets, since it involves the inversion of the Gram matrix.

Parameters:

list - the training list of samples

activeTrain

public void activeTrain(java.util.List<TrainingSample<T>> list,
                        int iterations,
                        int samples,
                        int pool)

projectSample

public double[] projectSample(T sample)

Project a sample to the space induced by the Nystrom approx

Parameters:

sample -

Returns:

projection of the sample

projectList

public java.util.List<TrainingSample<double[]>> projectList(java.util.List<TrainingSample<T>> l)

valueOf

public double valueOf(T t1,
                      T t2)

Description copied from class: Kernel

compute the kernel similarity between two element of input space

Specified by:

valueOf in class Kernel<T>

Parameters:

t1 - first element

t2 - second element

Returns:

the kernel value

valueOf

public double valueOf(T t1)

Description copied from class: Kernel

kernel similarity to zero

Specified by:

valueOf in class Kernel<T>

Parameters:

t1 - the element to compute the similarity to itself