Sign in

Berfin Simsek

@bsimsek.bsky.social
29 followers 41 following 3 posts

Deep Learning x {Symmetries, Structures, Randomness} 🦄 Researcher at Flatiron Computational Maths in NYC. PhD from EPFL. www.bsimsek.com

PostsRepliesMedia
Reposted by Berfin Simsek
Gautam Kamath @gautamkamath.com · 10/05/2025
It's been one year since Jim Simons passed away. He remains one of the greatest philanthropists of our time. His commitment to fundamental scientific research was truly extraordinary and remains essential for advancing humanity. www.simonsfoundation.org/2024/05/10/s...
simonsfoundation.org
Simons Foundation Co-Founder, Mathematician and Investor Jim Simons Dies at 86
Simons Foundation Co-Founder, Mathematician and Investor Jim Simons Dies at 86 on Simons Foundation
1297
Berfin Simsek @bsimsek.bsky.social · 03/05/2025
Below this threshold, a mild overparameterization (log k factor) for k index vectors is sufficient to match the neurons to the ideal vectors due to a coupon collector argument. 🤠
000
Berfin Simsek @bsimsek.bsky.social · 03/05/2025
When the ideal vectors form an equiangular frame, all learned weights converge to their average (no matter how much overparameterization is used), which has turned from a saddle to a local minimum after a certain threshold of the dot product. 😲
100
Berfin Simsek @bsimsek.bsky.social · 03/05/2025
Searching for an exact inverse map from learned weights to ideal (concept) vectors is an intricate geometry question, even for "simple" idealized models. 🧐
100
Reposted by Berfin Simsek
Terence Tao @teorth.bsky.social · 01/05/2025
DARPA's "Exponentiating mathematics" (expMath) program, which is launching a challenge to develop and evaluate "AI collaborators" for assist in decomposing and formalizing informal mathematical proofs, is now taking short abstract proposal submissions: sam.gov/opp/869c8d73...
sam.gov
SAM.gov
5256