PeopleofSuperintelligence

Researcher

Itay Safran

Itay Safran is a researcher in deep learning theory who proves mathematical results about neural networks, from the benefits of depth to the delayed generalization called grokking. He earned a master's degree and a doctorate at the Weizmann Institute of Science under Ohad Shamir, finishing in 2020, and with Shamir wrote Depth-Width Tradeoffs in Approximating Natural Functions with Neural Networks in 2017 and Spurious Local Minima are Common in Two-Layer ReLU Neural Networks in 2018. He then held postdoctoral posts at Princeton with Jason Lee and at Purdue with Paul Valiant before joining Ben-Gurion University of the Negev in 2024. In 2026, his paper To Grok Grokking: Provable Grokking in Ridge Regression, written with Mingyue Xu and Gal Vardi, received an Outstanding Paper Honorable Mention at ICML. He is a Senior Lecturer at Ben-Gurion University, affiliated with its Institute for the Foundations of AI.

Timeline

  1. 2015–2020

    Weizmann studies under Shamir

    He studies computer science and mathematics at Ben-Gurion University, then earns an M.Sc. and a Ph.D. at the Weizmann Institute of Science, advised by Ohad Shamir. He works on the doctorate from 2015 and finishes it in 2020.

  2. 2016–2018

    Depth and local minima at ICML

    With Ohad Shamir, he presents On the Quality of the Initial Basin in Overspecified Neural Networks at ICML 2016, Depth-Width Tradeoffs in Approximating Natural Functions with Neural Networks at ICML 2017, and Spurious Local Minima are Common in Two-Layer ReLU Neural Networks at ICML 2018. The 2017 paper proves that simple functions such as indicators of balls and ellipses are better approximated by deeper networks than by shallower ones, even if the shallower networks are much larger. The 2018 paper gives a computer-assisted proof of spurious local minima in two-layer ReLU networks, and experiments indicate that mild over-parameterization drastically reduces them.

  3. 2019–2021

    Depth separations and random shuffling

    In 2019, he coauthors Depth Separations in Neural Networks: What is Actually Being Separated? with Ronen Eldan and Ohad Shamir, which asks whether depth separations still hold for radial functions with a constant Lipschitz parameter. In 2021, he and Shamir prove in Random Shuffling Beats SGD Only After Many Epochs on Ill-Conditioned Problems that without-replacement SGD improves on with-replacement SGD in worst-case bounds only when the number of epochs exceeds the condition number.

  4. 2020–2022

    Princeton postdoc with Jason Lee

    He is a postdoctoral research fellow at Princeton University from 2020 to 2022, hosted by Jason D. Lee, with whom he writes Optimization-Based Separations for Neural Networks (COLT 2022). The paper proves that gradient descent can efficiently learn ball indicator functions with a depth 2 network whose hidden layer stays fixed during training. In 2022, he and Gal Vardi also coauthor with Lee On the Effective Number of Linear Regions in Shallow Univariate ReLU Networks, an analysis of gradient flow on univariate ReLU networks with a single hidden layer.

  5. 2022–2024

    Purdue postdoc with Paul Valiant

    He is a postdoctoral research associate at Purdue University from 2022 to 2024, hosted by Paul Valiant. With Daniel Reichman and Valiant, he writes How Many Neurons Does it Take to Approximate the Maximum? in 2023, which proves new lower and upper bounds on the width needed to approximate the maximum function at various depths.

  6. 2024

    Joining Ben-Gurion's faculty

    He joins the Faculty of Computer and Information Science at Ben-Gurion University of the Negev in 2024. His research focuses on theoretical machine learning, with emphasis on deep learning theory, and its links to approximation theory, optimization and theoretical computer science.

  7. 2026

    ICML 2026 honorable mention

    To Grok Grokking: Provable Grokking in Ridge Regression, which he coauthors with Mingyue Xu and Gal Vardi, receives an Outstanding Paper Honorable Mention at ICML 2026. The paper proves that gradient descent with weight decay on over-parameterized linear regression first overfits, then generalizes poorly for a long stretch, and finally reaches arbitrarily small generalization error.

  8. As of October 2026

    Senior Lecturer at Ben-Gurion

    He is a Senior Lecturer at Ben-Gurion University of the Negev, affiliated with its Institute for the Foundations of AI. In October 2026, he is an invited speaker in the artificial intelligence session of the university's national computer and information science conference, where he presents To Grok Grokking.

Career

  1. Since 2024Ben-Gurion University of the NegevSenior Lecturer
  2. 2022–2024Purdue UniversityPostdoctoral Research Associate
  3. 2020–2022Princeton UniversityPostdoctoral Research Fellow

Education

  1. 2020Weizmann Institute of ScienceM.Sc.
  2. Ben-Gurion University of the Negev

From the biography.

  1. The Massachusetts Institute of Technology trained or employed 44 of the 187 AI major-prize winners, more than any other universityIncludes Ben-Gurion University of the Negev. The Massachusetts Institute of Technology trained or employed 44 of the 187 AI major-prize winners, more than any other university. We compare doctorate and 5+ year employment links, career timing and current workplaces for 187 major-prize and 1,111 best-paper winners.
  2. More people shaping AI studied at Massachusetts Institute of Technology than at any other institutionIncludes Ben-Gurion University of the Negev. We estimate where the people shaping AI studied and where their careers led. Massachusetts Institute of Technology leads with 467 former students, including 291 recorded doctorates.
  3. The company share of AI researchers’ top papers was 31% in 2020–2024 and 8% in 1970–1974Includes Ben-Gurion University of the Negev. We estimate where the people shaping AI worked when their most-cited papers appeared: the company share was 31% in 2020–2024 and 8% in 1970–1974, and Google led all organizations in 2020–2024.
  4. 84% of people shaping AI with recorded undergraduate and doctoral schools changed institutions for their doctorateIncludes Princeton University. 84% of people shaping AI with recorded undergraduate and doctoral schools changed institutions for their doctorate. Across 2,116 people with both an undergraduate degree and a doctorate recorded, we trace which undergraduate schools feed which doctoral programs, how often students stay at the same school, and how paths differ by birth-country region.

Colleagues

People named in this biography, and people with dated positions at the same organization in the same years.

Sources and credits

Photo: We have no published photo of Itay Safran yet. Send us a free-licensed photo.

  1. Itay Safran | alphaXiv alphaxiv.org. Checked 21 October 2026.
  2. Students & Postdocs · Ohad Shamir cs.toronto.edu. Checked 21 October 2026.
  3. Itay Safran's Webpage - Publications sites.google.com. Checked 21 October 2026.
  4. Depth-Width Tradeoffs in Approximating Natural Functions with Neural Networks proceedings.mlr.press. Checked 21 October 2026.
  5. Spurious Local Minima are Common in Two-Layer ReLU Neural Networks proceedings.mlr.press. Checked 21 October 2026.
  6. [1904.06984] Depth Separations in Neural Networks: What is Actually Being Separated? arxiv.org. Checked 21 October 2026.
  7. [2106.06880] Random Shuffling Beats SGD Only After Many Epochs on Ill-Conditioned Problems arxiv.org. Checked 21 October 2026.
  8. [2112.02393] Optimization-Based Separations for Neural Networks arxiv.org. Checked 21 October 2026.
  9. [2205.09072] On the Effective Number of Linear Regions in Shallow Univariate ReLU Networks: Convergence Guarantees and Implicit Bias arxiv.org. Checked 21 October 2026.
  10. [2307.09212] How Many Neurons Does it Take to Approximate the Maximum? arxiv.org. Checked 21 October 2026.
  11. Itay Safran - Ben-Gurion University Research Portal cris.bgu.ac.il. Checked 21 October 2026.
  12. ICML 2026 Awards icml.cc. Checked 21 October 2026.
  13. [2601.19791] To Grok Grokking: Provable Grokking in Ridge Regression arxiv.org. Checked 21 October 2026.
  14. Artificial - National Conference on Computer and Information Science 2026, Ben-Gurion University conferences.bgu.ac.il. Checked 21 October 2026.