Speaker

Weiran Wang

Abstract

Multiple views of data, both naturally acquired and artificially produced, have been proven useful in representation learning. In this talk, I will discuss the weakly-supervised setup of multi-view learning and a series of methods based on Canonical correlation analysis (CCA) and the underlying probabilistic model, and focus on the question of whether we can disentangle the variations shared by both views, from the variations that are private to each view. We answer this question in the affirmative in two cases.

In the first case, we study CCA-type learning under an ideal data generation model, for which correlation maximization is guaranteed to extract the shared components across views, up to certain ambiguities. In addition, the private information in each view can be disentangled from the shared information using proper regularization design.

In the second case, we disentangle the static component from the dynamic component for sequence data. We derive a novel evidence lower bound for approximate likelihood estimation, involving mutual information (MI)-based regularizations that are key to disentanglement. We enhance contrastive estimation of MI with careful data augmentation to further assist separation of components.

Bio

Weiran Wang is currently a Staff Research Scientist at Google. He was a Senior Research Scientist at Salesforce Research from 2019 to 2020, and a Senior Applied Scientist at Amazon Alexa from 2017 to 2019. He spent 2014 to 2017 as a postdoc researcher at Toyota Technological Institute at Chicago, after obtaining PhD at University of California at Merced in 2013. His research interests lie in machine learning, speech processing, and optimization.