Anna Vasilevskaya
@loghyr.bsky.social
PhD student studying cortical computations in https://www.apredictiveprocessinglab.org
Most importantly, this framework is falsifiable – mapping JEPA architectures onto cortical circuits (see also the RPL proposal bsky.app/profile/fzen...) produces experimental predictions that we can test to further constrain the search space for algorithms of cortical function!
1/6 New preprint 🚀 How does the cortex learn to represent things and how they move without reconstructing sensory stimuli? We developed a circuit-centric recurrent predictive learning (RPL) model based on JEPAs.
🔗 doi.org/10.1101/2025...
Led by @atenagm.bsky.social @mshalvagal.bsky.social
🔗 doi.org/10.1101/2025...
Led by @atenagm.bsky.social @mshalvagal.bsky.social
January 30, 2026 at 2:37 PM
Most importantly, this framework is falsifiable – mapping JEPA architectures onto cortical circuits (see also the RPL proposal bsky.app/profile/fzen...) produces experimental predictions that we can test to further constrain the search space for algorithms of cortical function!
We think the idea of a JEPA could be an opportune starting point for developing the next iteration of a hypothesis for the algorithm of cortical function. It shows consistency with physiological and anatomical data and addresses many of the PP limitations.
January 30, 2026 at 2:37 PM
We think the idea of a JEPA could be an opportune starting point for developing the next iteration of a hypothesis for the algorithm of cortical function. It shows consistency with physiological and anatomical data and addresses many of the PP limitations.
This suggested to us that cortex might function similar to a JEPA @yann-lecun.bsky.social, in which deep and superficial cortical layers implement the encoder networks and prediction error is minimized in latent space rather than input space.
January 30, 2026 at 2:37 PM
This suggested to us that cortex might function similar to a JEPA @yann-lecun.bsky.social, in which deep and superficial cortical layers implement the encoder networks and prediction error is minimized in latent space rather than input space.
We investigated this idea further by developing a paradigm of artificial coupling between L5 activity and motor-related predictions in the absence of sensory input. We discovered that a mismatch between L5 activity and motor prediction is indeed sufficient to drive prediction error responses in L2/3
January 30, 2026 at 2:37 PM
We investigated this idea further by developing a paradigm of artificial coupling between L5 activity and motor-related predictions in the absence of sensory input. We discovered that a mismatch between L5 activity and motor prediction is indeed sufficient to drive prediction error responses in L2/3
However, we discovered that L5 functionally interacts with L2/3 like a bottom-up teaching signal, and we hypothesized that L2/3 might function to model and predict L5 activity rather than raw sensory input.
January 30, 2026 at 2:37 PM
However, we discovered that L5 functionally interacts with L2/3 like a bottom-up teaching signal, and we hypothesized that L2/3 might function to model and predict L5 activity rather than raw sensory input.
By probing the functional influence between neurons of deep and superficial cortical layers, we were able to test the experimental predictions the two implementation proposals make. Intriguingly, neither proposal could account for the interlaminar interactions that we observed!
January 30, 2026 at 2:37 PM
By probing the functional influence between neurons of deep and superficial cortical layers, we were able to test the experimental predictions the two implementation proposals make. Intriguingly, neither proposal could account for the interlaminar interactions that we observed!
Most prevalent circuit models of PP postulate laminar segregation between prediction error neurons and internal representation neurons. We set out to experimentally distinguish between two PP proposals that differ in their assumptions on whether the cortex is a hierarchy.
January 30, 2026 at 2:37 PM
Most prevalent circuit models of PP postulate laminar segregation between prediction error neurons and internal representation neurons. We set out to experimentally distinguish between two PP proposals that differ in their assumptions on whether the cortex is a hierarchy.
Reposted by Anna Vasilevskaya
100%, but also - the eLife experiment is a good reminer that we need experiments here. Open reviews were not a thing until they were, peer reviews were not a thing until they were, so did preprints, and the ability not to respond to reviewers comments etc. We can try to re-imagine publishing!
July 20, 2025 at 7:45 AM
100%, but also - the eLife experiment is a good reminer that we need experiments here. Open reviews were not a thing until they were, peer reviews were not a thing until they were, so did preprints, and the ability not to respond to reviewers comments etc. We can try to re-imagine publishing!