BISCUIT: Causal Representation Learning from Binary Interactions

Explore a comprehensive talk on causal representation learning from binary interactions, focusing on the BISCUIT method. Delve into the identification of causal variables in environments like robotics and embodied AI, where an agent's interactions can be described by unknown binary variables. Learn about the BISCUIT architecture, which simultaneously learns causal variables and their corresponding binary interaction variables. Examine experimental results from three robotic-inspired datasets, demonstrating BISCUIT's accuracy in identifying causal variables and its scalability to complex, realistic environments. Gain insights from the speaker, Phillip Lippe, as he discusses the method's applications, limitations, and potential future developments in the field of causal representation learning.

- Discussant Slide + Introduction
- BISCUIT Binary Interactions
- BISCUIT Architecture
- Experiments
- Conclusion
- Discussion