Proving and improving the Platonic Representation Hypothesis with JEPAs
The Platonic Representation Hypothesis is proven as a theorem, showing that JEPAs learn the same representation up to one rotation when trained on any number of modalities.
Key points
- •The Platonic Representation Hypothesis is proven as a theorem, providing guarantees and recipes.
- •JEPAs trained separately on any number of modalities learn the same representation up to one rotation.
- •Whether models trained on different modalities converge to the same representation is central to AI.
#Research & Papers#LLM / Model Trends#JEPAs