Colloquium Lecture: Kathleen Creel (University of Michigan), "Against the Platonic Representation Hypothesis"

Abstract: The Platonic Representation Hypothesis posits that “neural networks, trained with different objectives on different data and modalities, are converging to a shared statistical model of reality in their representation spaces” (Huh et al. 2024). This realist explanation for the success of contemporary large language and image models has gained traction within computer science, particularly as an explanation for some of the striking model convergence results found using mechanistic interpretability techniques. In this talk, I undermine the Platonic Representation Hypothesis by positing alternative explanations for the phenomenon of model convergence and propose a different test of convergence to replace it - one with its own roots in the history of philosophy.