Proving and Improving the Platonic Representation Hypothesis with JEPAs
30 points by Anon84
The study proves the Platonic Representation Hypothesis (PRH) by establishing a Platonic Representation Theorem, demonstrating that Joint-Embedding Predictive Architectures (JEPAs) trained on various modalities converge to a consistent representation of the world. This theorem confirms that as these models scale, they learn the same representation, differing only by a rotation. The findings provide a solid foundation for understanding how different AI models can align in their interpretations of the world.