A little late on the 🦋🚂, but happy to share a new(ish) preprint - the culmination of three years of postdoc with @tyrellturing.bsky.social and @apeyrache.bsky.social, and my first real foray into #NeuroAI as a tool to study the sleeping brain 🧠🟦: biorxiv.org/content/10.1... (1/🧵) c’est parti! - ThreadSky

dlevenstein.bsky.social • 207 days ago

A little late on the 🦋🚂, but happy to share a new(ish) preprint - the culmination of three years of postdoc with @tyrellturing.bsky.social and @apeyrache.bsky.social, and my first real foray into #NeuroAI as a tool to study the sleeping brain 🧠🟦:

https://biorxiv.org/content/10.1101/2024.04.28.591528v2

(1/🧵) c’est parti!

Comments

barbarikon.bsky.social•81 days ago

We published a replay-preplay based model of place field formation and navigation in Neural Networks last year.

A. Alabi, D. Vanderelst, A.A. Minai (2023) Neural Networks, 161: 116-128

Here's the arXiv version: https://arxiv.org/pdf/2206.02249

dlevenstein.bsky.social•81 days ago

Looks super interesting, thanks for sharing! 👀

barbarikon.bsky.social•81 days ago

Very nice to see the full paper. I enjoyed your poster at the NeuroAI workshop. Glad to see the citations of Chip Levy’s work. He was ahead of his time on this. I was his postdoc working on hippocampal sequence learning.

dlevenstein.bsky.social•207 days ago

The hippocampus is involved in diverse cognitive functions (navigation, memory, imagination, ...), which rely on the interplay between two modes of operation: an "online" mode when animals are engaged in active behavior and an "offline" mode during quiescent behavior and sleep.

dlevenstein.bsky.social•207 days ago

In the online mode, the hippocampus shows prominent theta rhythms, and sensory inputs drive a representation of space with spatially tuned cells and a neural manifold that maps the environment.

dlevenstein.bsky.social•207 days ago

In the offline mode, theta disappears and you see large irregular activity with sharp wave ripples. Rather than the animals current position, the hippocampus “replays” plausible trajectories through the environment and outputs learned associations for the rest of the brain.

dlevenstein.bsky.social•207 days ago

We currently lack a theory of how a network with this double life can emerge, and specifically how a network with access to only egocentric sensory inputs can develop a cognitive map that autonomously produces replay-like dynamics when that input is removed.

dlevenstein.bsky.social•207 days ago

One candidate is predictive learning. Classical theories called the hippocampus a prediction circuit (Levy 1989, Lisman/Redish 2009), its tuning curves resemble a predictive map (Stachenfeld et al 2017), and networks that predict vision develop place cells (Recanatesi et al 2021)

dlevenstein.bsky.social•207 days ago

However, it’s unclear if learning to predict produces a network that can intrinsically generate replay.
To test this, we trained neural networks in an environment in which visual input is spatially ambiguous, redundant, and comes from egocentric units with weak spatial tuning.

neuroai.bsky.social•193 days ago

Very cool paper! Look forward to seeing it published.

dlevenstein.bsky.social•193 days ago

Thanks Ida!🙏

Comments

Posting Rules

Reply