Abstract

Representation learning constructs low-dimensional representations to summarize essential features of high-dimensional data like images and texts. Ideally, such a representation should capture non-spurious features of the data in an efficient way. It shall also be disentangled so that we can freely manipulate each of its dimensions. However, these desiderata are often intuitively defined and challenging to quantify or enforce.

In this talk, we take on a causal perspective of representation learning. We show how desiderata of representation learning can be formalized using counterfactual notions, which then enables algorithms that target efficient, non-spurious, and disentangled representations of data. We discuss the theoretical underpinnings of the algorithm and illustrate its empirical performance in both supervised and unsupervised representation learning.

Speaker Bio:

I am an LSA collegiate fellow and an assistant professor of Statistics (as of Fall 2022) at the University of Michigan.

I work in the fields of Bayesian statistics, machine learning, and causal inference. I also work on algorithmic fairness and reinforcement learning. My research interests lie in the intersection of theory and applications.

Previously, I was a postdoctoral researcher with Professor Michael Jordan at the University of California, Berkeley. I completed my Ph.D. in Statistics at Columbia University, advised by Professor David Blei, and my B.Sc. in mathematics and computer science at the Hong Kong University of Science and Technology.