Kenji Fukumizu

Also known as 福水健次

Kenji Fukumizu is a statistical machine learning researcher at the Institute of Statistical Mathematics in Tokyo, whose work covers kernel methods, mathematical methods in deep learning and topological data analysis. After a mathematics PhD from Kyoto University in 1996 and time at Ricoh and RIKEN's Brain Science Institute, he joined the Institute of Statistical Mathematics in 2000 and became a professor there in 2009. With Francis Bach and Michael Jordan he published kernel-based dimensionality reduction methods in 2004 and 2009, and in 2010 he proposed the Kernel Bayes' rule with Le Song and Arthur Gretton. A Linear-Time Kernel Goodness-of-Fit Test, written with Wittawat Jitkrittum, Wenkai Xu, Zoltan Szabo and Gretton, won a NeurIPS Best Paper Award in 2017. He is a professor at the Institute of Statistical Mathematics and directs its Research Center for Statistical Machine Learning, a post he has held since 2012.

Now at: Research Organization of Information and Systems and Okinawa Institute of Science and Technology Graduate University
Badges: Bridge
Loyal

Timeline

1989–1996

Ricoh and a Kyoto PhD

In 1989, he earns a B.S. in mathematics at Kyoto University and begins work as a researcher in the research and development department of Ricoh. He completes a PhD in mathematics at Kyoto University's Graduate School of Science in 1996.
1998–2000

Moving from RIKEN to ISM

He joins RIKEN in January 1998 as a research scientist in the Lab for Information Synthesis of its Brain Science Institute. In February 2000, he becomes an associate professor at the Institute of Statistical Mathematics (ISM).
2002–2004

Kernel dimension reduction

From May 2002 to April 2003, he is a visiting scholar in Michael I. Jordan's lab at the University of California, Berkeley. In 2004, he publishes Dimensionality Reduction for Supervised Learning with Reproducing Kernel Hilbert Spaces with Francis R. Bach and Michael I. Jordan, a method that uses covariance operators on reproducing kernel Hilbert spaces to find a low-dimensional subspace that retains the statistical relationship between an explanatory variable and a response variable.
2009

Professor at ISM

He becomes a professor at the Institute of Statistical Mathematics in October 2009. The Annals of Statistics publishes his 2009 paper Kernel dimension reduction in regression, written with Francis R. Bach and Michael I. Jordan, which characterizes conditional independence through conditional covariance operators on reproducing kernel Hilbert spaces.
2010

The Kernel Bayes' rule

In 2010, he and coauthors Le Song and Arthur Gretton post Kernel Bayes' rule, a nonparametric method that realizes Bayes' rule through representations of probabilities in reproducing kernel Hilbert spaces. The posterior is expressed as the mean of a weighted sample in that space.
2017

A NeurIPS Best Paper Award

A Linear-Time Kernel Goodness-of-Fit Test, written with Wittawat Jitkrittum, Wenkai Xu, Zoltan Szabo and Arthur Gretton, receives a Best Paper Award at NeurIPS in 2017. The test learns features that best indicate differences between observed samples and a reference model, at a computational cost linear in the number of samples.
2018

Advising Preferred Networks

In 2018, he and Masaaki Imaizumi post Deep Neural Networks Learn Non-Smooth Functions Effectively, which derives generalization error rates for ReLU networks and shows the rates are almost optimal for estimating non-smooth functions. In November 2018, Preferred Networks hires him as a technical advisor to advance theoretical research on neural network models together with its researchers.
As of October 2026

Directing a machine learning center

He is a professor at the Institute of Statistical Mathematics in Tokyo and has directed its Research Center for Statistical Machine Learning since January 2012. His research covers mathematical methods in deep learning, topological data analysis and kernel methods, and he has been a visiting professor at the Okinawa Institute of Science and Technology since April 2023.