Current Position & Biography
Makoto Yamada is an associate professor at the Okinawa Institute of Science and Technology, where he leads the Machine Learning and Data Science Unit. His research connects statistical learning with explainable models, efficient distributed training and applications involving complex scientific data. Rather than concentrating on a single consumer AI product, he works on methods that can help researchers select informative variables, examine predictions and train models across multiple participants.
Yamada earned a bachelor's degree at the University of Aizu, a master's degree in electrical engineering at Colorado State University, and a doctorate in statistical science at SOKENDAI in 2010. His career includes roles at Tokyo Institute of Technology, NTT Communication Science Laboratories and Yahoo Labs. His dated institutional CV records an OIST associate professorship beginning in August 2022 and the end of his Kyoto associate-professor and RIKEN team-leader appointments in March 2023.
Notable Contributions
Yamada coauthored GraphLIME with Qiang Huang and colleagues. The research proposes a way to explain individual graph-neural-network predictions by identifying informative features in a node's local neighbourhood. It uses nonlinear feature selection to produce a local explanation. Its reported experiments support the method in the settings tested; an explanation is not itself proof that a prediction is causally correct or fair.
Another collaborative contribution is Block HSIC Lasso, published in Bioinformatics in 2019 with Héctor Climente-González, Chloé-Agathe Azencott and Samuel Kaski. The work investigates feature selection for very high-dimensional datasets and evaluates it on synthetic data and genomic datasets, including gene-expression, single-cell RNA-sequencing and genome-wide association data. It is a computational research method, not a clinically validated diagnostic product.
In 2026, Yamada, Kaoru Otsuka and Yuki Takezawa published Delayed Momentum Aggregation at ICML. Their DeMoA optimizer combines stored momentum from clients not sampled in a training round with fresh momentum from participating clients. The paper analyzes convergence and robustness against malicious or faulty participants under partial participation, and reports experimental benefits in the tested configurations. Such results do not establish immunity to every attack or automatically guarantee privacy.
Why Katharos Is Following Them
Yamada's work addresses practical questions behind dependable AI: what information influences a model, how distributed learning responds to unreliable participants, and which variables matter in scientific datasets. These questions connect statistical foundations to systems that must operate under communication, interpretability and security constraints.
Areas to Follow
Explainable graph models, high-dimensional feature selection, communication-efficient federated learning, robustness assumptions and reproducible evaluation across different datasets.
Reporting Scope
Coverage credits collaborators and distinguishes theoretical guarantees under stated assumptions from empirical performance or real-world deployment. Keeping raw data on local devices is not equivalent to a comprehensive privacy guarantee, and algorithmic research should not be presented as a medical treatment or a universal security solution.
