Sumio Watanabe

Overview

Sumio Watanabe is a mathematician and machine-learning theorist at the Tokyo Institute of Technology, the creator of singular learning theory — the use of algebraic geometry (resolution of singularities, the real log canonical threshold) to analyze the asymptotics of Bayesian learning in singular statistical models. He introduced the Widely Applicable Information Criterion (WAIC) and the widely applicable Bayesian information criterion (WBIC). (Distinct from Toshiaki Watanabe, the stochastic-volatility econometrician.)

Key Contributions / Features

• WAIC and its CV equivalence (Watanabe 2010): Proved the Widely Applicable Information Criterion is asymptotically equivalent to Bayes leave-one-out cross-validation for singular models, and that generalization + CV error $\to 2\lambda/n$. See Widely Applicable Information Criterion and Watanabe (2010).
• Singular learning theory (Watanabe 2001, 2009): Showed the asymptotic Bayes marginal likelihood of a singular model is governed by the real log canonical threshold, generalizing BIC; textbook Algebraic Geometry and Statistical Learning Theory (2009).
• WBIC (2013): A widely applicable Bayesian information criterion estimating the marginal likelihood for singular models from a single-temperature posterior.

Related

• Widely Applicable Information Criterion
• Bayesian Model Selection
• Watanabe (2010)