Alex Atanasov (@ABAtanasov) / X

Alex Atanasov

590 posts

Alex Atanasov

@ABAtanasov

Fascinated by scaling and universality. PhD from @harvardphysics. Not a Bayesian. Opinions my own 🇧🇬

New York

Joined January 2017

Alex Atanasov
@ABAtanasov
Apr 28, 2025
We went to the same (absurd) public HS and same grad school. Your take is spot on for our HS, but all of those kids became PMs at tiktok, traders at JS, and EMs at McKinsey. I've almost never seen this type go on to do quality research, and I've seen none of them do physics PhDs
sarah
@atheorist
Apr 27, 2025
Most people don't actually know the lengths parents will go to try to raise an academic superstar. In this post, I will detail the life of the average thoroughbred in STEM PhD programs at a top university. The thoroughbred lives a difficult life full of enormous amounts of
161K
Alex Atanasov
@ABAtanasov
Apr 28, 2025
Replying to @ABAtanasov
To do well at olympiads, get into college, and get the PM job just requires testing well and projecting confidence. There's no shortage of this in America. Doing research requires you to face being wrong/ignorant and humbles you in a way that usually breaks this kind of psyche.
8K
Alex Atanasov
@ABAtanasov
Apr 6, 2025
The mechanism behind double descent in ML (specifically in ridgeless least squares regression) is not just similar but _identical_ to that which in physics causes massless 4D phi^4 theory to go from being classically scale-free to picking up a scale/mass in the infrared.
40K
Alex Atanasov
@ABAtanasov
Apr 26, 2025
1/n I’m very excited to present this Spotlight. It was one of the more creative projects of my PhD, and also the last one with @blake__bordelon & @CPehlevan, the best coauthors you can have :) Come by this afternoon to learn "How Feature Learning Can Improve Neural Scaling Laws."
30K
Alex Atanasov
@ABAtanasov
May 3, 2024
[1/n] Thrilled that this project with @jzavatoneveth and @CPehlevan is finally out! Our group has spent a lot of time studying high dimensional regression and its connections to scaling laws. All our results follow easily from a single central theorem 🧵
arxiv.org
Scaling and renormalization in high-dimensional regression
From benign overfitting in overparameterized models to rich power-law scalings in performance, simple ridge regression displays surprising behaviors sometimes thought to be limited to deep neural...
22K
Alex Atanasov
@ABAtanasov
Apr 26, 2025
1/n Also excited to share this work at ICLR with Jamie Simon, @alexmeterez, and @CPehlevan . The energy on this project was electric – Alex and Jamie are true chefs at empirical deep learning. We're presenting on this at poster #616 this afternoon!
13K
Alex Atanasov
@ABAtanasov
Nov 4, 2021
(1/9) “Can neural networks at finite width still be kernel machines?” Very excited to announce the result of a collaboration with @blake__bordelon and @CPehlevan
arxiv.org
Neural Networks as Kernel Learners: The Silent Alignment Effect
Neural networks in the lazy training regime converge to kernel machines. Can neural networks in the rich feature learning regime learn a kernel machine with a data-dependent kernel? We demonstrate...
Alex Atanasov
@ABAtanasov
Aug 16, 2025
You can hate on string theory all you want but when you see results like this - that the only amplitudes compatible with the scattering constraints of quantum gravity are those of string theory - it really leaves you in awe of its universality. Any answer to “what is the theory
arxiv.org
Strings from Almost Nothing
We argue that string theory emerges inevitably from a few simple assumptions about physical scattering. Consistency alone requires that all tree-level four-point scattering amplitudes exhibit...
7.1K
Alex Atanasov
@ABAtanasov
Jan 8, 2023
1/11 Excited to share some recent work with @blake__bordelon, Sab Sainathan, and @CPehlevan on the effects of dataset, width, and initialization scale on neural network (NN) performance arxiv.org/abs/2212.12147
13K
Alex Atanasov
@ABAtanasov
Dec 8, 2023
Very happy to share this work in NeurIPS 2023 with @vyasnikhil96, @blake__bordelon, Sab Sainathan, @DepenKenpachi, and @CPehlevan on the consistent behavior of feature-learning networks across large widths arxiv.org/abs/2305.18411. What is large width consistency? Read on! 1/n
arxiv.org
Feature-Learning Networks Are Consistent Across Widths At Realistic Scales
We study the effect of width on the dynamics of feature-learning neural networks across a variety of architectures and datasets. Early in training, wide neural networks trained on online data have...
11K
Alex Atanasov
@ABAtanasov
Aug 8, 2025
If you’re an undergrad interested in shaping the future of AI, you can either fine-tune underwhelming language models at a leading AI lab OR you can join the group of one of the most talented scientists I know, pursuing some of the most promising research at the intersection of
Blake Bordelon ☕️🧪👨‍💻
@blake__bordelon
Aug 8, 2025
Excited to announce that I will be joining @UTAustin with a joint position between @OdenInstitute for Computational Science and dept of Neuroscience in FL 2026! I plan on recruiting PhD students and postdocs interested in mathematics of neural computation (more details to come).
14K
Alex Atanasov
@ABAtanasov
Oct 8, 2024
Incredible news for our field!! It’s so personally meaningful to see deep learning get recognized as a scientific field that builds on top of the physics canon
Nobel Prize in Physics 2024
From nobelprize.org
2.9K
Alex Atanasov
@ABAtanasov
Apr 7, 2025
Replying to @nblqbl
We do our best to bridge the ML-physics gap here: arxiv.org/abs/2405.00592 though there are many other ML-oriented reviews I can also recommend, such as this one stat.berkeley.edu/~ryantibs/stat…
arxiv.org
Scaling and renormalization in high-dimensional regression
From benign overfitting in overparameterized models to rich power-law scalings in performance, simple ridge regression displays surprising behaviors sometimes thought to be limited to deep neural...
993
Alex Atanasov
@ABAtanasov
Jan 21, 2023
Really happy that this effort was accepted to ICLR 2023 :)
Alex Atanasov
@ABAtanasov
Jan 8, 2023
1/11 Excited to share some recent work with @blake__bordelon, Sab Sainathan, and @CPehlevan on the effects of dataset, width, and initialization scale on neural network (NN) performance arxiv.org/abs/2212.12147
4.1K