7/7 I've attempted a schematic summarizing JEPA, NeoSSL, and RPL, plus their mapping to the canonical microcircuit. It’s 4am here 😅, so I hope these are helpful clarifications!

6/7 I view RPL as an alternative circuit implementation of JEPA, rather than the singular introduction of it. That said, the successor representation analysis and PFC abstraction results are excellent, another strong signal for SSL in cortical circuits 💪

5/7 L2/3->L5 connection constitutes the Predictor (similar to BYOL). Moreover, the ffw from L2/3 to L5 is good mapping to W_{l23->L5} (synapses from L2/3 to L5) and it's just a linear layer without any non-linearity (you can set it as identity matrix if you wish)

4/7 Even with AE regularization, L5 is jointly trained via the SSL prediction loss. Its representation is modulated by predictability, not just compression. Framing L2/3 as merely supervised prediction of a fixed AE latent state is factually incorrect. Regardless, still a JEPA!

3/7 Crucially, our Supp. Mat. (Fig S5) shows that NeoSSL works equally well using VICReg (variance maximization) instead of reconstruction to avoid collapse. The specific regularization method is interchangeable; the architecture remains a JEPA regardless.

2/7 LeCun (2022) defines JEPA by prediction in latent space, not by how collapse is avoided. NeoSSL maps pixels to lower-dim embeddings & minimizes latent distance (Joint). How one prevents collapse (criterion 2) shouldn't be conflated with the architecture itself.

1/7 Great work on RPL! :) I’m a big fan of SSL in the cortex. I just want to clarify a few points regarding the comments and potential confusion about our NeoSSL model (Nejad et al. 2025).

Android is still buggy on X/Twitter. Elon should have used an android phone for a week to feel the pain of Android users. Although, I'm going to miss bookmarks. that was the only twitter feature I used a lot.