Neural Variational Random Field Learning

Volodymyr Kuleshov, Stefano Ermon

Abstract: We propose variational bounds on the log-likelihood of an undirected probabilistic graphical model p that are parametrized by flexible approximating distributions q. These bounds are tight when q = p, are convex in the parameters of q for interesting classes of q, and may be further parametrized by an arbitrarily complex neural network. When optimized jointly over q and p, our bounds enable us to accurately track the partition function during learning.

Conflicts: mcgill.ca