PeopleofSuperintelligence

Builder

Roy Frostig

Roy Frostig is a machine-learning researcher and JAX co-creator whose work connects learning theory with numerical computing. He studied in Stanford's Machine Learning Group with Percy Liang and coauthored a 2013 semantic-parsing paper. He created JAX with colleagues and coauthored an early tracing paper, a study of data-parallel training, and papers on automatic differentiation. In 2024, he coauthored a study of supervised learning from independent trajectories and the Gemini 1.5 report; in 2025, he coauthored the Gemini 2.5 report. He is a research scientist at Google DeepMind and a core developer of JAX.

Works
29
Citations
2,907
h-index
13
Badges
Links
Website, Google Scholar, X and GitHub

Publication numbers from OpenAlex, as of 22 October 2026.

Timeline

  1. 2012–2013

    Stanford PhD and semantic parsing

    He joins Stanford's Machine Learning Group as a PhD student in 2012, with Percy Liang as his advisor. In 2013, he coauthors “Semantic Parsing on Freebase from Question-Answer Pairs” with Jonathan Berant, Andrew Chou and Liang; the paper appears at EMNLP.

  2. 2015

    A one-pass estimator

    In 2015, he coauthors “Competing with the Empirical Risk Minimizer in a Single Pass” with Rong Ge, Sham M. Kakade and Aaron Sidford at COLT. The paper seeks estimators that match the empirical risk minimizer’s statistical rate on every problem, using one pass, linear time and space proportional to one sample.

  3. 2017–2018

    Creating and compiling JAX

    He creates JAX with a few colleagues in 2017. In 2018, he coauthors “Compiling machine learning programs via high-level tracing” with Matthew Johnson and Chris Leary; the paper describes a tracing JIT compiler that generates accelerator code from pure Python and Numpy programs through XLA.

  4. 2018

    Batch size and training time

    In 2018, he coauthors “Measuring the Effects of Data Parallelism on Neural Network Training” with Christopher Shallue, Jaehoon Lee, Joseph Antognini, Jascha Sohl-Dickstein and George E. Dahl. The study measures how increasing batch size changes the steps and time needed to reach a target error.

  5. 2021–2022

    Automatic differentiation papers

    In 2021, he coauthors “Efficient and Modular Implicit Differentiation” with Mathieu Blondel, Marco Cuturi, Jean-Philippe Vert and others; it differentiates optimization problems through a Python function that states their optimality conditions. In 2022, he coauthors “You Only Linearize Once: Tangents Transpose to Gradients” with Alexey Radul, Adam Paszke, Matthew Johnson and Dougal Maclaurin; the paper decomposes reverse-mode differentiation into forward-mode differentiation followed by transposition of its linear part.

  6. 2024

    Learning from many trajectories

    In 2024, he coauthors “Learning from many trajectories” with Stephen Tu and Mahdi Soltanolkotabi. The paper studies supervised learning from independent sequences of dependent covariates, including settings in sequence modeling, control and reinforcement learning.

  7. 2024–2025

    Gemini reports

    In 2024, he coauthors the Gemini 1.5 report, which describes multimodal models that recall and reason over fine-grained information from millions of tokens of context. In 2025, he is among the authors of the Gemini 2.5 report, which introduces Gemini 2.5 Pro and Gemini 2.5 Flash.

  8. As of October 2026

    Research scientist at Google DeepMind

    As of October 2026, he is a research scientist at Google DeepMind and a co-author and core developer of JAX. His work covers machine-learning foundations through theory, systems and programming languages.

Career

  1. NowGoogle DeepMindResearch Scientist
  2. 2018–2023Google

Education

  1. Stanford UniversityPhD

From the biography and OpenAlex affiliations (the years of papers that list the organization), so some years are approximate.

Most-cited works

  1. Semantic Parsing on Freebase from Question-Answer Pairs2013 · Conference on Empirical Methods in Natural Language Processing (EMNLP) · 1,580 citations
  2. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context2024 · 302 citationsWith George van den Driessche and Paul Barham
  3. Measuring the Effects of Data Parallelism on Neural Network Training2018 · 230 citations
  4. Toward Deeper Understanding of Neural Networks: The Power of Initialization and a Dual View on Expressivity2016 · 190 citationsWith Yoram Singer
  5. Compiling machine learning programs via high-level tracing2018 · 177 citationsWith Chris Leary

From OpenAlex, as of 21 October 2026.

  1. The median h-index is 48 at universities and 14 at frontier AI labsIncludes Roy Frostig. The median h-index is 48 at universities and 14 at frontier AI labs. We compare publishing careers, recent citations, dual appointments and job moves across 2,420 university researchers and 308 frontier-lab researchers.
  2. Google had the most people shaping AI every year from 2014 to 2025Includes Google. Our estimates of which organizations employed the most people shaping AI each year from 1955 to 2025, tracing the shift from university labs to companies.
  3. Google has the highest combined h-index among AI organizationsIncludes Google. We estimate that Google holds the highest combined h-index at 8,016, while universities account for 73% of the deduplicated total across AI organizations.
  4. 20% of people at companies who shape AI also hold a university role, led by Google and Stanford UniversityIncludes Google. We estimate that 539 of the 6,766 people with a known current organization (8%) hold both a university role and a company role, including 20% of people at companies and 13% of people at universities.

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 Roy Frostig yet. Send us a free-licensed photo.

  1. Machine Learning Group statsml.stanford.edu. Checked 21 October 2026.
  2. Semantic Parsing on Freebase from Question-Answer Pairs - ACL Anthology aclanthology.org. Checked 21 October 2026.
  3. Competing with the Empirical Risk Minimizer in a Single Pass proceedings.mlr.press. Checked 21 October 2026.
  4. Roy Frostig cs.stanford.edu. Checked 21 October 2026.
  5. Compiling machine learning programs via high-level tracing research.google. Checked 21 October 2026.
  6. Measuring the Effects of Data Parallelism on Neural Network Training arxiv.org. Checked 21 October 2026.
  7. Efficient and Modular Implicit Differentiation arxiv.org. Checked 21 October 2026.
  8. You Only Linearize Once: Tangents Transpose to Gradients arxiv.org. Checked 21 October 2026.
  9. Learning from many trajectories jmlr.org. Checked 21 October 2026.
  10. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context arxiv.org. Checked 21 October 2026.
  11. Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities. arxiv.org. Checked 21 October 2026.