A Unifying Principle for the Functional Organization of Visual Cortex biorxiv.org/cgi/content/sh… #biorxiv_neursci

Very excited to finally share this work! More details in thread: x.com/eshedmargalit/…

Getting the chance to work on this project w/ Eshed Margalit has been one of the most inspiring & exciting parts of my PhD. Check out this paper, which represents many years of clever, careful & thought-provoking work from Eshed - not to mention it’s an engineering tour-de-force!

And what's striking to me is how detailed and fine-grained the predictions about the loss circuitry can be -- I'm beginning to feel that we can really expect goal-driven end-to-end trained neural network models to yield fairly specific circuit insights.

Excited to share this work unveiling how maps emerge in visual cortex congrats to Eshed who worked fearlessly for 4+ years to solve this difficult problem!

Machine vision is a mess of special methods, Frankensteined together. For awhile I thought it had to be - like the evolutionary goulash of our own visual system. But no. Vision can be UNIFIED with a framework we call COUNTERFACTUAL WORLD MODELING (CWM). arxiv.org/abs/2306.01828

Code for playing with Counterfactual World Models is now available: github.com/neuroailab/Cou… Some fun things you can do (with your own images or ours):

We at @noetik_ai are thrilled to introduce the *first world model of spatial, multimodal cancer biology*: Oncology Counterfactual Therapeutics Oracle (OCTO) Technical report below + 🧶 on how we trained OCTO directly on patient data and use it to find new cancer drug targets!

We NOETIK built a new foundation model for spatial biology: OCTO-VirtualCell. This model is prompted to simulate single cell gene expression in "virtual cells," which are placed within real, intact patient tissue. Read about & explore these simulations below: