A Phenomenology of Neural Representation (Thomas Fel) - Interpretable Deep Learning Seminars

Interpretable Deep Learning Seminar Series Speaker: Thomas Fel (Goodfire) Title: A Phenomenology of Neural Representation Abstract Debates that shaped systems neuroscience (single neurons, population codes, manifolds) are quietly replaying inside vision, vision-actions and large language models, yet interpretability has stayed largely method-driven. In this talk, we argue for a phenomenology of neural representation: treating geometric structure as the object of study, clarifying what current representation hypotheses actually claim, and motivating a new wave of interpretability built from the phenomenon up. Interpretable Deep Learning website: https://interpretabledeeplearning.git...