Scholar
Matus Telgarsky
Google Scholar ID: Fc-5yRIAAAAJ
Courant Institute of Mathematical Sciences, New York University
deep learning theory
machine learning theory
Follow
Homepage
↗
Google Scholar
↗
Citations & Impact
All-time
Citations
5,597
H-index
23
i10-index
33
Publications
20
Co-authors
25
list available
Contact
GitHub
Open ↗
Publications
3 items
Understanding Reasoning from Pretraining to Post-Training
2026
Cited
0
Benefits of Early Stopping in Gradient Descent for Overparameterized Logistic Regression
2025
Cited
0
Convex Analysis at Infinity: An Introduction to Astral Space
arXiv.org · 2022
Cited
10
Resume
Academic Achievements
2013: Showed coordinate descent (and steepest descent) converges to maximum margin solutions
2016: Constructed deep networks inapproximable by shallow ones unless exponentially wide
2018: Extended classical margin-based generalization theory to deep networks
2019: Large-margin analysis of gradient descent for non-separable data, with a succinct SGD proof (1/t rate)
2020: Proved directional convergence of gradient descent for shallow and deep ReLU networks near initialization
2024: Interpreted (decoder-)Transformer layers as parallel computation rounds; showed logistic regression is insensitive to step size
Co-authors
17 total
Daniel Hsu
Columbia University
Peter Bartlett
Professor, EECS and Statistics, UC Berkeley
Sham M Kakade
Harvard University
Anima Anandkumar
California Institute of Technology and NVIDIA
Dylan J. Foster
Principal Researcher, Microsoft Research
Clayton Sanford
Google Research
Maxim Raginsky
Professor of Electrical and Computer Engineering, University of Illinois, Urbana-Champaign
Alexander Rakhlin
Professor, MIT