Hossein Taheri

I am a machine learning researcher, interested in different problems arising in deep learning including efficient optimization, fine-tuning, alignment, robustness and generalization capabailites of neural nets and LLMs. Previouly I was a postdoctoral researcher in the department of Computer Science and Engineering at UC San Diego and obtained my Ph.D. in Electrical and Computer Engineering from UC Santa Barbara. I completed my B.Sc. in Electrical Engineering and Mathematics at Sharif University of Technology.

Research Interests

• Machine Learning and Optimization Algorithms
• Deep Learning Theory: Generalization and Optimization of Neural Networks
• Transformers and In-context Learning
• Continual Learning
• Distributed Optimziation
• High-dimensional Statistics

Publications

Ph.D. Thesis:

1. Generalization and Optimization in the Interpolation Regime: From Linear Models to Neural Networks [link]
   H. Taheri, University of California, Santa Barbara, 2024 .

   Peer-Reviewed Publications:

   2. On the Theory of Continual Learning with Gradient Descent for Neural Networks [slides] [link] [Openreview] [code]
      H. Taheri, A. Ghosh, A. Mazumdar (ICML 2026).
   3. Sharper Guarantees for Learning Neural Network Classifiers with Gradient Methods [code] [link] [poster] [Openreview]
      H. Taheri, C. Thrampoulidis, A. Mazumdar (ICLR 2025).
   4. On the Optimization and Generalization of Multi-head Attention [link]
      P. Deora*, R. Ghaderi*, H. Taheri*, C. Thrampoulidis (* denotes equal contribution) (TMLR 2024 and selected for poster presentation at ICLR 2025).
   5. Generalization and Stability of Interpolating Neural Networks with Minimal Width [link][slides]
      H. Taheri, C. Thrampoulidis (JMLR 2024).
   6. On Generalization of Decentralized Learning with Separable Data [link]
      H. Taheri, C. Thrampoulidis (AISTATS 2023).
   7. Fast Convergence in Learning Two-layer Neural Networks with Separable Data [code][link]
      H. Taheri, C. Thrampoulidis (AAAI 2023).
   8. Asymptotic Behavior of Adversarial Training in Binary Linear Classification [code][link]
      H. Taheri, R. Pedarsani, C. Thrampoulidis (IEEE TNNLS 2023).
   9. Fundamental Limits of Ridge-Regularized Empirical Risk Minimization in High Dimensions [code] [link]
      H. Taheri, R. Pedarsani, C. Thrampoulidis (AISTATS 2021).
   10. Quantized Decentralized Stochastic Learning over Directed Graphs [code][link]
       H. Taheri, A. Mokhtari, H. Hassani, R. Pedarsani (ICML 2020).
   11. Sharp Asymptotics and Optimal Performance for Inference in Binary Models [link]
       H. Taheri, R. Pedarsani, C. Thrampoulidis (AISTATS 2020).
   12. Robust and Communication-efficient Collaborative Learning [code] [link]
       A. Reisizadeh, H. Taheri, A. Mokhtari, H. Hassani, R. Pedarsani (NeurIPS 2019).

   Other Publications:

   13. Asymptotic Behavior of Adversarial Training in Binary Linear Classification[link]
       H. Taheri, R. Pedarsani, C. Thrampoulidis (ISIT 2022).
   14. Sharp Guarantees and Optimal Performance for Inference in Binary and Gaussian-mixture Models[link]
       H. Taheri, R. Pedarsani, C. Thrampoulidis (Entropy, 2021).
   15. Optimality of Least-squares for Classification in Gaussian Mixture Models[link]
       H. Taheri, R. Pedarsani, C. Thrampoulidis (ISIT 2020).
   16. Sharp Guarantees for Solving Random Equations with One-bit Information[link]
       H. Taheri, R. Pedarsani, C. Thrampoulidis (Allerton 2019).