Maybe we should write a joint perspective on this to clarify the differences and similarities :)

Thanks both! L5 encoder can learn in different ways (as we show). We can do EMA as in JEPA. Also the L2/3>L5 input is not important. Both these points were only added for biological plausibility. The fundamental principle of JEPA is the same: a predictor (L2/3) learns to match a target encoder (L5).

Cool work! I should clarify that our cortical SSL model (Nejad 2025) also predicts in the latent space like JEPA. Indeed the mapping to JEPA is relative straigtforward (L4 and L5 are encoders and L2/3 is the predictive network).

Led by @dabalpedamonti.bsky.social @mohintasamia.bsky.social in collaboration with Hugo Malagon-Vina and Stephane Ciocchi.

Thanks for this beautiful description of our paper!