CMPS 290c S07: Final Project: MCMC Algorithms for solving Multitask Co-Leraning Problem

This project presents the di¢ culties in bounding the mixing time of the Markov chain presented in [1] for solving the multitask co-learning problem. In particular I show why the technics of canonical paths and coupling can not be directly applied. Then I propose an alternative markov chain that approximate the notion of co-learning for which a bounding in the mixing time is possible. The upper bound, unfortunately, is not polynomial in m -the number of ’good’ learners. To try to cope with that I give possible alternatives that might allow a bounding that depend only polynomialy in m.