Heading to Anaheim for #APSSummit25! I'll be giving a talk at GSNP's MAR-G54 session.
summit.aps.org/events/MAR-G54

Looking forward to connecting with old and new colleagues #APS2025 @apsphysics.bsky.social

Intriguingly we found that for models trained in a carefully controlled setting (fixed data diet, architecture, and number of epochs), there is very little objective-function-related spread in neural predictivity across different cortical regions.

13/n

Instead MMCR optimizes a population level feature (the spectrum) directly.
 
Training to optimize the capacity of augmentation manifolds produces a representation that linearly separates semantic classes of images at a level that is on par with popular SSL frameworks

10/n

Specifically, we maximize the nuclear norm of a matrix containing view-averaged unit norm embeddings: Batch of images --> multiple synthetic views per image --> embed on unit sphere --> average over views --> maximize nuclear norm of a matrix containing the centroids.

8/n

We introduce a new self-supervised learning method called MMCR (maximum manifold capacity representation), inspired by methods from statistical physics that use geometry to characterize the coding capacity of a learned representation.

Let's dig in!
2/n

Excited to share our results on Efficient Coding of Natural Images using Maximum Manifold Capacity Representations, a collaboration with Teddy Yerxa, Yilun Kuang, Eero Simoncelli to be presented at #NeurIPS2023

🧵1/n

@tedyerxa.bsky.social
@flatironinstitute.org @simonsfoundation.org