Scientific Computation and Applied & Industrial Mathematics Seminar: Alexandre Bouchard-Cote

Computational biology, spatio-temporal analysis, natural language processing and a range of other fields rely on increasingly complex probabilistic models to make predictions and take action. In practice, these models often need to incorporate high-dimensional latent variables, complex combinatorial spaces and various heterogeneous data-structures. Moreover, it is important to not only perform optimization on these models, but also to assess the uncertainty in predictions and reconstructions of latent states.

Monte Carlo methods have been used in the past several decades to approach these hard and important problems. Advances in probabilistic programming open the door for more widespread use of Monte Carlo, but computational efficiency remains a formidable challenge.

In this talk, I will provide some background on the state-of-the-art, and describe the progress that my collaborators and myself have been making towards more practical Monte Carlo methods. In particular, I describe Divide-and-Conquer Sequential Monte Carlo (D&C SMC), a method for performing inference on a collection of auxiliary distributions organized into a tree. In contrast to standard SMC samplers, D&C SMC exploits multiple populations of weighted particles, while still being an exact approximate method. D&C SMC is applicable to a broad class of probabilistic graphical models, including models with loops.