Homepage

Welcome to my homepage. My name is Behrad Moniri (بهراد منیری) and I am a Ph.D. student working on the theory of machine learning at the University of Pennsylvania where I am advised by Prof. Hamed Hassani. My research interests include deep learning theory, random matrix theory, and information theory.

In particular, I am interested in studying exactly-solvable models of deep learning. I propose and analyze these models and use them to explore the emerging behaviors of modern machine learning methods.

Here you can find my CV. I can be reached at bemoniri@seas.upenn.edu, and on Twitter @bemoniri.

Education

Doctor of Philosophy (Ph.D.), Electrical and Systems Engineering
School of Engineering and Applied Sciences, University of Pennsylvania, 2021 – present.

Master of Arts (A.M.), Statistics and Data Science
Wharton School of Business, University of Pennsylvania, 2021-2024.

Bachelor of Science (B.Sc.), Electrical Engineering
Sharif University of Technology, 2016-2020.

News

• Our feature learning paper is accepted to ICML 2024.

• I have been awarded the Outstanding Teaching Award by the School of Engineering and Applied Sciences, University of Pennsylvania, April 2024.
  

• I will be presenting our feature learning results at the workshop honoring Prof. Andrew Barron: Forty Years at the Interplay of Information Theory, Probability and Statistical Learning, Yale Institute for Foundations of Data Science, April 26 to 28, 2024.

• I will be giving an invited talk at the INFORMS Optimization Society (IOS) Conference, Session on “Nonconvex Optimization Approaches in Machine Learning”, March 2024.

• I am invited to give a talk at the workshop “Beyond classical regimes in statistical inference and machine learning”, Centre International de Math ́ematiques et Informatique de Toulouse (CIMI), Nov. 2024. (can not attend due to visa issues)

Publications

Google Scholar, DBLP

(* equal contributions)

• A Theory of Non-Linear Feature Learning with One Gradient Step in Two-Layer Neural Networks
  Behrad Moniri*, Donghwan Lee*, Hamed Hassani, Edgar Dobriban.
  International Conference on Machine Learning (ICML 2024).
  [ArXiv] 
  
  
• Demystifying Disagreement-on-the-Line in High Dimensions
  Donghwan Lee*, Behrad Moniri*, Xinmeng Huang, Edgar Dobriban, Hamed Hassani.
  International Conference on Machine Learning (ICML 2023).
  [ArXiv] [PMLR] [Poster]
  
  
• Information-Theoretic Analysis of Minimax Excess Risk
  Hassan Hafez-Kolahi*, Behrad Moniri*, Shohreh Kasaei.
  IEEE Transactions on Information Theory, 2023.
  [ArXiv] [IEEE]
  

• Rate-Distortion Analysis of Minimum Excess Risk in Bayesian Learning
  Hassan Hafez-Kolahi, Behrad Moniri, Shohreh Kasaei, and Mahdieh Soleymani.
  International Conference on Machine Learning (ICML 2021). [Oral Presentation]
  [ArXiv] [PMLR] [Poster]

Slides

• A Theory of Non-Linear Feature Learning in Two-Layer Neural Networks [Slides]
  
• Demystifying Disagreement-on-the-Line in High Dimensions [Slides]
  
• Asymptotic Risk of High-Dimensional Linear Regression [Slides]
  
• High-Dimensional Regression: From Double Descent to Distribution Shifts [Slides]
  
• Precise Tradeoffs in Adversarial Training in Linear Regression [with S. Hadou] [Slides]  
  
• Universal Inference [Slides]
  
  For some older slides, please see here.