O Speaker: Dr. Jongha Ryu, postdoc associate at MIT
O Title: Generative Modeling with Succinct Representation via Wyner’s Common Information
O Date: Jan. 18 (Wednesday)
O Start Time: 11 AM
O Venue: N1, Room 201
In this talk, I will introduce an information-theoretic approach to develop a deep generative model for two correlated high-dimensional random vectors such as images and captions, for solving machine learning tasks such as joint and conditional sample generation and cross-domain retrieval. Based on Wyner’s common information from network information theory, I will motivate a new bimodal generative model framework that seeks a succinct common representation, and describe how to train the proposed model systematically.