9 November 2021

Jan-Willem van de Meent

Compositional Inference in Probabilistic Programs

Compositional Inference in Probabilistic Programs


Speaker: Jan-Willem van de Meent, Associate Professor (UHD), AMLab, University of Amsterdam


Deep probabilistic programming systems combine the principles of deep learning with the principles of probabilistic modeling. The user programmatically specifies a deep generative model (a neural mapping from latent variables to data), along with a corresponding inference model (a neural mapping from data to latent variables), which together can be trained using stochastic gradient descent with little or no supervision.

In this talk, I will discuss recent innovations in training deep probabilistic programs by combining techniques from variational inference and importance sampling. For many years, deep generative models were typically trained by maximizing a reparameterized lower bound, as is done in variational autoencoders. However, this approach can fail to converge to a meaningful representation in more structured problems, such as tasks the involve reasoning about shared features for a small batch of inputs. I will discuss how we can overcome these difficulties, using variational methods that learn proposals for importance samplers, as well as a programming abstractions for high-level specification of such methods in probabilistic programming systems.

Watch Back ›