ogpstep_ep

Purpose

Updates the TAP/EP parameters after an online sweep

Synopsis

ogpstep_sp(is_sp,indX,K1,K2,gamma,hatE,tA,tV,logZ);

Description

ogpstep_ep(is_sp,indX,K1,K2,gamma,hatE,tA,tV,logZ) - updates the GLOBAL structure ep, the parameters of the TAP/EP approximation to reflect the addition of the current input.

This function is called from ogppost and is not used on itw own.

Depending on the indicator is_sp, the updates in the structure ep correspond to a full/sparse OGP update.

Parameters

net,ep

- the GLOBAL Gaussian Process and TAP/EP structures which, although not in the list of parameters, all operations are performed on these structures.

isSp

- a boolean value which tells the procedure which typeof update -- sparse/non-sparse -- is carried out.

indX

- the index of the current input in the training set (needed to update the TAP/EP representation).

kX

- the covariance function of the basis vector set and the current input K(BV,x).

K1,K2

- the values of the first and second derivatives of the log-average.

gamma

- the distance of the current input from the basis vector set.

hatE

- the coordinates of the projection of the current input to the basis vector set.

tA,tV

- the values of the cavity means and variances - obtained from the TAP/EP pre-adjustment step.

logZ

- the local approximation to the evidence, computed using the cavity field and the likelihood function.

See Also

ogp, ogptrain, ogpstep_full, ogpstep_sp


Pages: Index

Copyright (c) Lehel Csató (2001-2004)