CARE: BISCUIT - Causal Representation Learning from Binary Interactions

Abstract

I was invited to give a talk on our recent work BISCUIT. BISCUIT is a method for learning the causal variables of an interative environment that is observed from high-dimensional inputs such as images. Using low-level action information, BISCUIT can be applied to many practical environments, such as for embodied AI. In this talk, I gave a detailed overview of the theory, the architecture, and our experimental results. For more details, check out our paper and project page.