Deep Mixture and Generative Models: Representation, Approximation, Robustness, and Application

Recent progress on deep architectures has enabled efficient representation and learning of complex high dimensional probability distributions over rich sensory data. In particular, deep mixture models and deep generative models have emerged as the most powerful techniques for this task. The proposed research aims at addressing some of the fundamental questions in this field: What is the relationship between the two seemingly different methodologies? What is the trade-off between network size and noise complexity, and how do these factors affect the approximation accuracy? Can we build on recent advances on sum of squares relaxations to design principled and provably correct convex relaxations for parameter estimation? And last but not the least, how do we measure and strengthen the robustness of such unsupervised deep models? TO BE CONT’D

Faculty Supervisor:

Yaoliang Yu

Student:

Partner:

Royal Bank of Canada (Borealis)

Discipline:

Computer science

Sector:

Information and Communications Technology; Technology; Finance and Insurance

University:

University of Waterloo

Program: