Jascha Sohl-Dickstein

Career

Jascha Sohl-Dickstein worked as a research associate on NASA Jet Propulsion Laboratory's Mars Exploration Rovers from 2001 to 2005 and completed a PhD in biophysics at UC Berkeley in 2012. Sohl-Dickstein was then a visiting scholar in applied physics at Stanford and, from 2012 to 2014, worked at Khan Academy, an online education nonprofit.¹

Sohl-Dickstein joined Google Brain in 2015 as a senior research scientist and became a principal scientist at Google DeepMind in 2023. In 2024 Sohl-Dickstein joined Anthropic as a Member of Technical Staff.¹²

Research

Sohl-Dickstein's 2015 paper with Eric Weiss, Niru Maheswaranathan and Surya Ganguli introduced diffusion models, which learn to reverse a gradual noising process. A 2021 follow-up with Yang Song and others framed score-based generative modeling with stochastic differential equations.¹³⁴

Sohl-Dickstein's CV also groups work on the theory of infinite-width neural networks, showing that such networks are linear in their parameters throughout training and that their outputs are Gaussian processes; on learned optimizers, including VeLO (2022); and on adversarial examples that influence both machine and human perception.¹

Open-source work

Sohl-Dickstein co-led BIG-bench, a collaborative benchmark of diverse tasks for probing large language models and extrapolating their capabilities, and wrote its original design document. The CV also lists Neural Tangents, a library for defining, training and evaluating both finite and infinite-width neural networks.¹⁵

Work and contributions

• Diffusion models (Deep Unsupervised Learning using Nonequilibrium Thermodynamics) Published research

  First author, with Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli, of the ICML 2015 paper that Sohl-Dickstein's CV lists as the original paper introducing diffusion models; also a co-author of the 2020 paper on score-based generative modeling through stochastic differential equations.³¹⁴

• BIG-bench (Beyond the Imitation Game benchmark) Open-source code

  Co-led the project and wrote the original design document, per Sohl-Dickstein's CV; the google/BIG-bench repository is licensed under the Apache License 2.0.¹⁵

Timeline

1. 2001Began work as a research associate on the Mars Exploration Rovers at NASA's Jet Propulsion Laboratory.¹
2. May 2012Completed a PhD in biophysics at UC Berkeley.¹
3. 2015Joined Google Brain as a senior research scientist.¹
4. Mar 2015Posted "Deep Unsupervised Learning using Nonequilibrium Thermodynamics," introducing diffusion models.³
5. Nov 2020Co-authored "Score-Based Generative Modeling through Stochastic Differential Equations."⁴
6. 2023Became a principal scientist at Google DeepMind.¹
7. 2024Joined Anthropic as a Member of Technical Staff.¹

Other documented roles

Roles the sources document that are not confirmed as current.

• Principal Scientist (2023–2024); earlier Senior Staff Research Scientist, Staff Research Scientist, and Senior Research Scientist (2015–2023), Google Brain and Google DeepMind¹²
• Visiting Scholar (2012–2015), Stanford University (Applied Physics)¹

Honors and recognition

• ICML Distinguished Paper Award for "Minimum probability flow learning" · International Conference on Machine Learning, 2011¹
• ICLR Outstanding Paper Award for "Score-Based Generative Modeling through Stochastic Differential Equations" · International Conference on Learning Representations, 2021¹
• ICML Outstanding Paper Award for "Unbiased Gradient Estimation in Unrolled Computation Graphs with Persistent Evolution Strategies" · International Conference on Machine Learning, 2021¹

Where they work

• Anthropic 34 people in the wiki

Related people

• Diederik P. KingmaAlso at Anthropic
• Kai FronsdalAlso at Anthropic
• Collin BurnsAlso at Anthropic
• Samuel R. BowmanAlso at Anthropic
• Jack ClarkAlso at Anthropic
• Stefano ErmonImage, video, and 3D generation · Machine-learning theory
• Richard ZhangImage, video, and 3D generation · Evaluation and benchmarks
• Yang SongImage, video, and 3D generation · Machine-learning theory
• Michael CollinsMachine-learning theory · Evaluation and benchmarks
• Ruchir PuriEvaluation and benchmarks