jsat.math.optimization.stochastic

Class NAdaGrad

java.lang.Object

jsat.math.optimization.stochastic.NAdaGrad

All Implemented Interfaces:

java.io.Serializable, GradientUpdater

public class NAdaGrad
extends java.lang.Object
implements GradientUpdater

Normalized AdaGrad provides an adaptive learning rate for each individual feature, and is mostly scale invariant to the data distribution. NAdaGrad is meant for online stochastic learning where the update is obtained from one dataum at a time, and it relies on the gradient being a scalar multiplication of the training data. If the gradient given us a ScaledVector, where the base vector is the datum, then NAdaGrad will work. If not the case, NAdaGrad will degenerate into something similar to normal AdaGrad.

The current implementation assumes that the bias term is always scaled correctly, and does normal AdaGrad on it.
See: Ross, S., Mineiro, P., & Langford, J. (2013). Normalized online learning. In Twenty-Ninth Conference on Uncertainty in Artificial Intelligence. Retrieved from here

See Also:

Serialized Form

Constructor Summary

Constructors

Constructor and Description
NAdaGrad() Creates a new NAdaGrad updater
NAdaGrad(NAdaGrad toCopy) Copy constructor

Method Summary

Modifier and Type	Method and Description
NAdaGrad	clone()
void	setup(int d) Sets up this updater to update a weight vector of dimension d by a gradient of the same dimension
void	update(Vec w, Vec grad, double eta) Updates the weight vector x such that x = x-ηf(grad), where f(grad) is some function on the gradient that effectively returns a new vector.
double	update(Vec w, Vec grad, double eta, double bias, double biasGrad) Updates the weight vector x such that x = x-ηf(grad), where f(grad) is some function on the gradient that effectively returns a new vector.

Methods inherited from class java.lang.Object

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

Constructor Detail

NAdaGrad

public NAdaGrad()

Creates a new NAdaGrad updater

NAdaGrad

public NAdaGrad(NAdaGrad toCopy)

Copy constructor

Parameters:

toCopy - the object to copy

Method Detail

update

public void update(Vec w,
                   Vec grad,
                   double eta)

Description copied from interface: GradientUpdater

Updates the weight vector x such that x = x-ηf(grad), where f(grad) is some function on the gradient that effectively returns a new vector. It is not necessary for the internal implementation to ever explicitly form any of these objects, so long as x is mutated to have the correct result.

Specified by:

update in interface GradientUpdater

Parameters:

w - the vector to mutate such that is has been updated by the gradient

grad - the gradient to update the weight vector x from

eta - the learning rate to apply

update

public double update(Vec w,
                     Vec grad,
                     double eta,
                     double bias,
                     double biasGrad)

Description copied from interface: GradientUpdater

Updates the weight vector x such that x = x-ηf(grad), where f(grad) is some function on the gradient that effectively returns a new vector. It is not necessary for the internal implementation to ever explicitly form any of these objects, so long as x is mutated to have the correct result.

This version of the update method includes two extra parameters to make it easer to use when a scalar bias term is also used

Specified by:

update in interface GradientUpdater

Parameters:

w - the vector to mutate such that is has been updated by the gradient

grad - the gradient to update the weight vector x from

eta - the learning rate to apply

bias - the bias term of the vector

biasGrad - the gradient for the bias term

Returns:

the value to change the bias by, the update being bias = bias - returnValue

clone

public NAdaGrad clone()

Specified by:

clone in interface GradientUpdater

Overrides:

clone in class java.lang.Object

setup

public void setup(int d)

Description copied from interface: GradientUpdater

Sets up this updater to update a weight vector of dimension d by a gradient of the same dimension

Specified by:

setup in interface GradientUpdater

Parameters:

d - the dimension of the weight vector that will be updated