| |
Kernel dependency estimation - for general input-output relations
|
This version does an eigendecomposition of the output kernel L
and learns targets by learning each orthogonal direction independently.
The pre-image algorithm is just to choose closest pre-image from
training set.
A=KDE(H) returns a kde object initialized with hyperparameters H.
Hyperparameters, and their defaults
feat=[] -- number of KPCA features to train on,
default take eigenvalues with lambda>lambda_max/10000
ridge=1e-5 -- regularization on input kernel
child=kernel -- the kernel for inputs stored as member "child"
ok=kernel -- the kernel for outputs
output_preimage=0 -- output index from training sample of preimage
instead of actual label
use_pca=1 -- can remove decorrelation part for quick & dirty
approximation (also allows to handle non-Mercer
kernels)
Model
alpha -- the weights
Xsv -- the Support Vectors
uv -- decomposition in output space
Methods:
train, test
| |
Reference : Kernel Dependency Estimation |
| |
Author : Jason Weston, Olivier Chapelle, Andre Elisseeff, Bernhard Schoelkopf and Vladimir Vapnik |