Neural Adaptive Smoothing via Twisting

Michael Y. Li, Dieterich Lawson, Scott Linderman

Keywords: inference, sequential latent variables, sequential monte carlo, discrete latent variables, reweighted wake sleep, variational inference

TL;DR: We combine smoothing sequential Monte Carlo and reweighted wake sleep to derive a method for model learning and inference in sequential latent variable models.

Abstract: We propose NAS-X, a method for sequential latent variable model learning and inference that uses smoothing sequential Monte Carlo (SMC) in a reweighted wake sleep (RWS) framework. Our method works with both discrete and continuous latent variables, and successfully fits a wider range of models than filtering SMC-based methods. We evaluate NAS-T on several tasks and find that it substantially outperforms existing methods in both inference and parameter recovery.