Online update coeffs. for binary classification.
[loglik,K1,K2] = c_class_bin(likpar,y,mx,varx);
[loglik,K1,K2] = c_class_bin(likpar,y,mx,varx,pM)
- returns the coefficients
for the online update of the GP assuming classification.
Parameters:
likpar
- the likelihood parameter (output noise).
y
- observed noisy labels.
mx
- mean of the GP marginal at the new input.
varx
- the variance of the GP marginal at the new input.
loglik
- the logarithm of the averaged likelihood function.
K1,K2
- the first and second derivatives of loglik
.
Although in practice not important, the noise variance is net.likpar(1) which should be set to a nonzero value when initialising the model.
The likelihood function is
Phi(x) = int(exp(-t^2/2),-infty,x+pM)/sqrt(2*pi)
which has the advantage of providing analytic formulae for the log-average.
(see matlab code for details).
ogp
, ogptrain
, err_2class
, demogp_class
Copyright (c) Lehel Csató (2001-2004)