Alternative Inductive Principles for Learning Restricted Boltzmann
Machines
By Ben Marlin
I'll give a short introduction to Boltzmann machines and restricted
Boltzmann machines. I'll discuss maximum likelihood learning in RBMs, and
focus on several alternatives including contrastive divergence,
pseudo-likelihood, ratio matching, and generalized score matching. I'll
present a comparison of the resulting learning algorithms on several data
sets and tasks including log likelihood evaluation, classification,
de-noising, and novelty detection. I'll also be showing a live, interactive
demo of all of the methods on a hand written digit recognition task.
Everyone is welcome to try out the demo following the talk.
Visit the LCI Forum page