exciting news 🚨

stevehanneke and I have finally nailed down the sharp realizable SVM margin-based generalization bound

"Empirical PAC-Bayes bounds for Markov chains" written with my PhD student Vahe Karagulyan accepted for a spotlight presentation at AISTATS 🎉

Looking forward to traveling to Tangier :)

arxiv.org/abs/2509.20985

new paper out

arxiv.org/abs/2601.04079

my return appearance on @DeGatchi's podcast!!

creators.spotify.com/pod/profile/...

to appear in ICML 2025:

The Empirical Mean is Minimax Optimal for Local Glivenko-Cantelli

with Doron Cohen and Roi Weiss

it's basically this paper, with a slightly changed title

arxiv.org/abs/2410.02835

check out this awesome new journal!

always a pleasure seeing my work built upon and generalized

I'm interested in something like a variable-input-length autoencoder, which maps vectors of varying dimension into a fixed dimension, with the usual objective of minimizing distortion (say, under the dynamic time warp distance, which handles variable-length arguments)

any pointers much welcome

I asked a friend to present a non-trivial information-theoretic inequality to ChatGPT-o1

I am pleased to report that the result was complete rubbish

anyone else have better luck with this?

see discussion here:

x.com/aryehazan/st...

just assigned this as an exercise

wish them luck

what is known about this exponent as a function of Q -- and in particular, its entropy? I think I can give a simple lower bound in the finite-alphabet case but curious about the SOTA

@mraginsky.bsky.social @roydanroy.bsky.social @thegautamkamath @ccanonne.bsky.social @pierrealquier.bsky.social

heck out my interview with
@DeGatchi
on
@ScrapingBits
!!

it was a lot of fun

creators.spotify.com/pod/show/scr...

math people: if I were to start augmenting my theorem proving with AI, which engine should I use?

willing to splurge grant money like a drunken sailor

hi I'm new on here