Regularization Paths for Learning Multiple
Kernels Description The follow-path package is a Matlab program that implements the Path
Following algorithm for Multiple Kernel Learning. Many regression and
classification problems that make use of a kernel can be extended so as to
simultaneously optimize the kernel itself. We call these Multiple Kernel
Learning problems. The Path Following algorithm returns all solutions of such a
problem as the intensity of the regularization varies. This set of solutions
forms a piecewise-smooth path, which the algorithm follows using
predictor-corrector steps. For more information, please read the
following paper: Precise instructions on how to
use the package are included in the archive file. In short, to use it you will
call follow_entire_path(), supplying it with the kernel matrices you want
to use and the training output, as well as a loss function (common examples are
provided but you may create your own). All other relevant parameters have widely
applicable defaults. The package contains a readme file with more detailed
instructions, and all user functions have an associated matlab help. Demo scripts are
also included. The follow-path package is Copyright (c) 2004, 2006 by Francis Bach and Romain Thibaux. If
you have any questions or comments regarding this package, or if you want to
report any bugs, please send an e-mail to
francis.bach@mines.org. The current
version 1.1 has been released on October 18th 2006. Check this web page regularly
for newer versions. follow-path
version 1.0 We would like to use inverse Hessian updating to reduce the complexity of each step of the path.
Demos
Last updated: October 18th, 2006
Matlab
code - version 1.1
Future
improvements
Demos
Francis R. Bach, Romain Thibaux, Michael I. Jordan.
Computing regularization paths for
learning multiple kernels.
Advances in Neural Information Processing Systems (NIPS) 17, 2005. [pdf]
follow-path
version 1.1 (NEW: unbalanced classification)
In our paper
we provide results for various real and synthetic regression and classification
problems. Each of these examples is provided in the package so that all
illustrations in the paper can be easily reproduced. This includes the very
simple 2D example used in the paper to illustrate the geometry. Detailed results can be found in the
paper.