Graduate Student Seminar - Bogdan Mazoure
This seminar will feature Bogdan Mazoure, who will tell us about variational auto-encoders.
Abstract:
In many statistical inference scenarios, it is computationally inefficient or even impossible to calculate the true solution to a given problem. Instead, we select a class of approximate distributions (for example, Normal) and minimize the divergence between the true and approximate solutions through maximization of a lower bound. Variational Bayesian methods are based on the free energy minimization principle used to explain physical systems and have a number of advantages over traditional inference models. In this talk, we will go over the main idea of approximate inference and discuss a neat use of variational inference in deep learning via variational auto-encoders.