Mathieu Blondel

Mathieu Blondel is a research scientist at Google DeepMind in Paris who studies differentiable programming, the end-to-end differentiation of computer programs so that their parameters can be tuned by gradient-based optimization. He coauthored the 2011 paper describing scikit-learn, a Python library he helped build as a core developer, and earned a PhD in machine learning from Kobe University in 2013. From 2013 to 2019, he was a researcher at NTT Communication Science Laboratories in Kyoto; in 2017, he and Marco Cuturi introduced Soft-DTW, a differentiable loss for time series. After moving to Google's Brain team in Paris in 2019, he was lead developer of the JAXopt optimization library and, with Vincent Roulet, released a draft of the book The Elements of Differentiable Programming in March 2024. His recent work links autoregressive language models with energy-based models, in a paper published at ICML in 2026.

Now at: Google DeepMind

Timeline

2011

Core developer of scikit-learn

In 2011, he coauthors the Journal of Machine Learning Research paper describing scikit-learn, a Python module of machine learning algorithms, with Fabian Pedregosa, Gaël Varoquaux and others. He serves as a core developer of the library, which later lists him among its emeritus maintainers.
2013

A PhD from Kobe University

He earns a PhD in machine learning from Kobe University in Japan in 2013. His May 2013 paper in the journal Machine Learning on block coordinate descent for sparse multiclass classification, written with Kazuhiro Seki and Kuniaki Uehara, receives an open-science honorable mention at ECML PKDD 2013.
2013–2019

Researcher at NTT in Kyoto

From 2013 to 2019, he is a researcher at NTT Communication Science Laboratories in Kyoto, Japan. With Akinori Fujino and Naonori Ueda, he publishes convex factorization machines at ECML PKDD in September 2015 and higher-order factorization machines at NIPS in December 2016.
2017

Soft-DTW with Marco Cuturi

In 2017, he and Marco Cuturi publish Soft-DTW at ICML, a differentiable loss function for time series built on a smoothed form of dynamic time warping. The loss suits averaging and clustering time series and training models whose outputs are time series.
2019–2020

Fenchel-Young losses

With André F. T. Martins and Vlad Niculae, he introduces Fenchel-Young losses in January 2019, a generic way to build a convex loss function for a regularized prediction function. The framework unifies many well-known losses, and the paper appears in volume 21 of the Journal of Machine Learning Research in 2020.
2019–2021

Google Brain in Paris

In 2019, he moves to Google in Paris as a research scientist on the Brain team, and by June 2021 he is a senior research scientist there. With Olivier Teboul, Quentin Berthet and Josip Djolonga, he presents fast differentiable sorting and ranking at ICML in July 2020. In 2021, he and seven coauthors propose automatic implicit differentiation of optimization problems, and he serves as lead developer of JAXopt, a JAX library of hardware-accelerated, batchable and differentiable optimizers.
March 2024–June 2025

A book on differentiable programming

In March 2024, he and Vincent Roulet release a draft of the book The Elements of Differentiable Programming, which reviews the end-to-end differentiation of computer programs from the perspectives of optimization and probability. In June 2025, he teaches a differentiable programming course based on the book at MLSS Senegal.
As of October 2026

Research scientist at Google DeepMind

He is a research scientist at Google DeepMind in Paris and contributes to gemma, a JAX library for using and finetuning Gemma models. At ICML in July 2026, he and four coauthors establish an explicit bijection between autoregressive language models and energy-based models. He co-supervises PhD student Germain Vivier-Ardisson with Axel Parmentier.