Convex Relaxations of Latent Variable Training (M44)
Yuhong Guo, Dale Schuurmans University of Alberta Problem Training in presence of hidden variables Use EM? Problem: local maxima

Solution: convex relaxation of joint (Viterbi) EM Difficulty: symmetry proper ty  trivial results if convex objective depends on hidden variable values directly Solution: reformulate the objective using equivalence relation matrices: Mij = 1 iff yi = yj for missing variable y Promising results on synthetic and real world Bayesian networks