This directory contains the matlab codes for running multi-class RVM, as in the tech report:

@techreport{Zhang:EECS-2005-6,
    Author = {Hao Zhang and Jitendra Malik},
    Title = {Selecting Shape Features Using Multi-class Relevance Vector Machine},
    Institution = {EECS Department, University of California, Berkeley},
    Year = {2005},
    Month = {October 10},
    Number = {UCB/EECS-2005-6},
    url={http://www.eecs.berkeley.edu/Pubs/TechRpts/2005/EECS-2005-6.pdf}
}


The main functions is reestimate.m, and a test on a toy case in test.m. 

The structure of the code is as follows:

outer loop:  reestimate_phi.m (reestimate_X.m)
inner loop:  logist_phi.m (logist_X.m) 

The inner loop uses a Newton-Raphson descent which is abstracted in newton.m

NOTES:
(1) For 2-class problems where the features are not class dependent (cf. Appendix A.1 in the tech report), a shortcut in computation is implemented in reestimate_X.m. 


(2) Direct your comments/suggestions to nhz@eecs.berkeley.edu