Victor Veitch: Linear Structure of (Causal) Concepts in Generative AI

- Title: Linear Structure of High-Level Concepts in Text-Controlled Generative Models, and the role of Causality
- Speaker: Victor Veitch (University of Chicago)
- Discussant: Francesco Locatello (Institute of Science and Technology Austria)
- Abstract: Text controlled generative models (such as large language models or text-to-image diffusion models) operate by embedding natural language into a vector representation, then using this representation to sample from the model's output space. This talk concerns how high-level semantics are encoded in the algebraic structure of representations. In particular, we look at the idea that such representations are ``linear''---what this means, why such structure emerges, and how it can be used for precision understanding and control of generative models.