Hawes Hall, Classroom 201, Harvard Business School
Title: Towards causally motivated robustness to distribution shifts
Abstract: One of the common failure modes of machine learning (ML) models is their lack of robustness to distribution shifts. In this talk, I will present solutions to this problem inspired by causal literature.
I will focus on a major failure mode that contributes to lack of robustness under distribution shifts: shortcut learning. Shortcut learning is a phenomenon in which models learn unstable associations that break as the test distribution shifts. I will present my work on developing causally-motivated regularization techniques that discourage shortcuts by encouraging models to conform to a pre-specified causal structure. I will discuss settings where possible shortcuts are known a priori and settings where such knowledge is not available. I will also discuss settings where the data is measured with some error.
