Japanese b Home b NTT CS Labs.

Biography and Summary of Research Achievements


Naonori Ueda respectively received B.S., M.S., and PhD degrees in Communication Engineering from Osaka University, Osaka, Japan in 1982, 1984, and 1992. In 1984, he joined the Electrical Communication Laboratories of NTT in Japan, where he researched image processing, pattern recognition, and computer vision. In 1991, he joined the NTT Communication Science Laboratories. From 1993 to 1994, he was a visiting scholar at Purdue University, West Lafayette, USA. He was the director of the NTT Communication Science Laboratories from April 2010 to March 2013. He is currently the head of the Ueda Research Laboratory (NTT Fellow) and the director of the Machine Learning and Data Science Center. He also serves as the deputy director of the RIKEN Center for the Advanced Intelligence Project, which was established in April, 2016. His current research interests include statistical machine learning and its applications to such natural sciences as medical, seismology, and meteorology. He is a visiting professor at the following three institutions: the Graduate School of Informatics, Kyoto University, Kobe University, and Meiji University. In 2019, he also serves as the research supervisor of Mathematical Information Platform, Japan Science and Technology Agency (JST). He is a member of the Information Processing Society of Japan (IPSJ), a fellow of the Institute of Electronics, Information, and Communication Engineers in Japan (IEICE), and a senior member of IEEE.

Summary of Research Achievements

In the early 1990s, as multi-layer neural networks were falling out of fashion, research had just started on statistical machine learning based on mathematical statistics, which can be described as the progenitor of modern machine learning. This is when he first started doing basic research in the field of machine learning. He also worked on a number of crucial cutting-edge projects, including Bayesian learning theory and its application to speech recognition. He has made significant technical contributions in the field of statistical machine learning, such as Nonparametric Bayes theory and its applications to relational data analysis. He worked as a sub-leader at the Funding Program for World-Leading Innovative R&D on Science and Technology (FIRST), at which he is a central research representative, while simultaneously designing an automatic technique for recognizing actual nursing activities from acceleration sensor data. Using this technique, he became the first scientist to analyze some nine million actions, an achievement that would not have been possible with conventional technology. Next, he launched a project called the gspatio-temporal multidimensional collective data analysis researchh as a representative of the NTT Machine Learning Data Science Center (NTT MLC) and developed a spatio-temporal prediction method applicable to unstationary data and real-time proactive people flow navigation technology. As the director of the Goal-oriented Technology Research Group at RIKEN AIP, he has been managing science research and social issues such as disaster prevention in Japan,@such as disaster prevention and mitigation, through machine learning technologies.

The following are his primary research achievements.
(1) Basic machine learning methods:
Naonori Ueda improved the quality of the solutions obtained by vector quantization, which is a critical technique used in the distortion-tolerant compression of such media as speech and images. He also discovered novel conditions necessary for obtaining optimal solutions and devised an online learning algorithm that implements these design principles by approximation. This research used new principles and algorithms to solve the issue of optimal quantizer design, resulting in a major step forward in the optimization of such quantizer design technology. For this work, he received an award from the Telecommunications Advancement Foundation in 1997. He is currently involved in a pioneering study of current online clustering.

Ueda also devised a novel method called Deterministic Annealing Expectation Maximization (DAEM) by applying the concepts of statistical mechanics to the poor local optimal problem associated the EM algorithm and devised a method called Split-Merge Expectation Maximization (SMEM) for solving local optimality problems in mixed models. Both techniques greatly improved the quality of parameter estimation solutions over a wide range of applications and gained international acclaim, including an introduction in an eminent mathematical statistics textbook.

Ueda rapidly became one of Japan's authorities on advanced statistical machine learning, such as variational Bayesian estimation as an approximate calculation method for large-scale statistical models. He also collaborated with researchers in speech recognition to develop the world's first model structure automatic learning method using a hidden Markov model, for which his research team received several research awards. His research group developed novel algorithms for relational data analysis based on nonparametric Bayes theory for handling infinite dimensional data partitioning. Uedafs team also developed methods for spatio-temporal statistical data analysis, including spatio-temporal point process using general intensity functions with deep neural networks.

(2) Applications based on machine learning approach:
In conventional pattern classification, a single pattern is assumed to be attributable to a unique class. When a single document consists of multiple classes like the text data on the web, conventional pattern classification techniques are inapplicable. To address this multiple classification issue, Ueda devised the first multiplex topic text model (PMM: Parametric Mixture Model) and demonstrated its usefulness in tests where it performed multiplex topic classification of tens of thousands of actual web pages. This technique was put into practical use as the topic classification engine of the news article classification service at the NTT Group portal site (Goo). This multiplex classification problem was later presented at an international workshop, where he joined the administration committee as a pioneering researcher.

Although techniques for behavior recognition from diverse sensors have recently been studied both in Japan and overseas, Uedafs work on the automatic behavior recognition of 41 different nursing actions from four acceleration sensors constitutes a unique contribution in this field. In this problem where conventional technology achieved a recognition rate of no more than 30%, he proposed a meta-learning method that remarkably boosted the success rate to over 60%. In collaboration with medical researchers at FIRST, this technique was used to analyze about nine million actions by nurses in the cardiac ward of Saiseikai Kumamoto Hospital. This collaboration is the first time useful statistics have been gathered on the relationships among nursing actions, the time taken to perform them, and the severity of the patients' conditions. This research introduced an entirely new avenue of big data analysis in ICT applications to healthcare.

As the director of NTT MLC, he launched spatio-temporal multidimensional collective data analysis research and developed a novel real-time, proactive navigation approach based on spatio-temporal predictions and gwhat-ifh simulations. An NTT operating company used this software in its actual operations. This series of spatio-temporal multidimensional collective data analysis studies was also commended by the Asia Pacific Signal Processing Society.

As the team leader of the Disaster Resilience Science Team at RIKEN AIP, Ueda has been managed novel machine learning approaches for several important research themes in the field of natural hazards and seismology. Specific topics include earthquake damage estimation, earthquake occurrence forecasting, and landslide susceptibility mapping. In 2019, our research results received the SC2017 Best Poster Award and the Best Paper Award from the International Society for Photogrammetry and Remote Sensing.

Association Memberships




Japanese b Home b NTT CS Labs.