Why Neural Network Representations Won’t Converge to Reality

Why we picked it

**This is a ticketed event. Please register at [this link](https://luma.com/n5d2q50u).** ​The Platonic Representation Hypothesis (Huh et al., 2024) proposes that neural networks are converging toward a shared statistical model of objective reality, offering a Platonic explanation for the growing representational similarity observed across models of different architectures, datasets, and modalities. In this talk, Robert Adragna argues that such convergence is computationally intractable in practice. Truly representing reality would require models to recognize the same real-world concept across all its possible appearances — a form of robustness that theoretical and empirical work suggests is infeasible to achieve at scale. He proposes an alternative account: representational convergence reflects shared structural assumptions embedded in training data, not the discovery of objective reality. ​​​**Event Schedule** 6:00 to 6:30 - Food and introductions 6:30 to 7:30 - Presentation and Q&