Connectionists: nonlinear dimensionality reduction

[Sayan Mukherjee]

Dear colleagues:

We have been developing in a series of papers
the use of gradient estimates for variable
selection, feature selection, as well as nonlinear
dimensionality reduction. All in the supervised
setting. This is related to the work in statistics
on Sliced Inverse Regression (SIR) and Sliced
Average Variance Estimation (SAVE) but for
high-dimensional problems. It is also related to
work on learning manifolds. An extension
to diffusion maps is in progress.

The papers are

Learning gradients and feature selection on manifolds:
http://ftp.stat.duke.edu/WorkingPapers/06-20.pdf

Estimation of Gradients and Coordinate Covariation in Classification
http://jmlr.csail.mit.edu/papers/v7/mukherjee06b.html

Learning Coordinate Covariances via Gradients
http://jmlr.csail.mit.edu/papers/v7/mukherjee06a.html

Matlab code can be downloaded at
http://www.stat.duke.edu/~sayan/soft.html


cheers,
sayan

-- 
Dr. Sayan Mukherjee				phone: 919 668-4747
Assistant Professor				fax:   919 668-0795
Computational Biology, Statistics, 		www.stat.duke.edu/~sayan
& Computer Science
Duke Scholar in Genomic Medicine
Duke University