Excited to share our latest story! We found disentangled memory representations in the hippocampus that generalized across time and environments, despite the seemingly random drift and remapping of single cells. This code enabled the transfer of prior knowledge to solve new tasks

Intelligent behavior involves generalizing from experience and applying prior knowledge to new situations. Traditional models attribute this to gradual learning in the cortex. However, such a slow process can't account for animals’ rapid generalization from limited experience

We found that the hippocampus supports rapid generalization in mice by generating disentangled memory representations, where different aspects of experience are encoded independently. These low-dimensional population dynamics evolved with learning.

We show that place cell representational drift and remapping are not random at the population level and supported by dedicated circuit mechanisms. This enabled the generalization of task variables (task structure, choice, time) across environments.

We believe that these findings advance a new framework to explain how the hippocampus achieves rapid generalization while also preserving the specificity of individual experiences. Check out the preprint for many more details:

This was only possible due to the work of an amazing team: first authors @hongyu_chang and Wenbo Tang, with @canliu18 , Salma Perez, Willian Zheng, Jaehyo Park. And of course Azahara Oliva 😍

Match Summary 💙 Get an absolute overview of your server several times - with the new Match Summary. Relax and never miss a game again. Activate the new function for your server today. Try Soccer Bot™:

@hongyu_chang Wow! I wrote a paper showing that disentangled representation learning + associative memory achieves zero-shot generalization in vision-based RL: Very much inspired by the idea that the hippocampus might be doing this. Cool to see real evidence for it!

@hongyu_chang Amazing work! Interested in " representational drift/remapping are not random... and supported by dedicated circuit mechanisms... generalization of task variables..." Working on plasticity that might depend on circuit states Do you think plasticity is directed & needs algorithms?

@hongyu_chang Manifolds. Manifolds everywhere.

@hongyu_chang I come from a deep learning background and this sounds so interesting. I'm able to grasp some of what you are saying, but are you able to provide a really intuitive high level description? What does it mean to have disentangled representations in these context, mathematically?

@hongyu_chang Incredible! excited to dive into this research