Research
Major Project Involvement
Applications Accepted (FY2023)
| No. of applications accepted | Amount (in thousands of yen) |
| 24 | 1,121,716 |
*Large-scale projects: Grants-in-Aid for Scientific Research S or higher grade, ERATO, CREST,PRESTO, MIRAI, and other projects with an annual research budget of ¥20 million or more.
Japan Science and Technology Agency (JST) CREST: Trusted AI Systems
Machine Learning That Connects to Symbolic Reasoning
With the success of deep learning, AI technology is being applied in wide-ranging areas of study, from the life sciences, physics, and chemistry to social and economic fields. However, as deep learning and other machine learning technologies come to be implemented in the real world and used in various situations, we are realizing issues inherent to machine learning, such as lack of interpretability, vulnerability to extrapolation, and biases, as well as issues that have emerged in relation to surrounding areas. It is vital to take these issues seriously and develop trustworthy AI. Aiming to develop the fundamental technologies that this will be based on, this research project will bring together modern machine learning using massive parameters, with symbolic reasoning offering high interpretability of the reasoning process. We focus mainly on a geometric approach and design and create a machine learning system based on symbolic reasoning, which simultaneously addresses the issues of reliability of machine learning and the robustness of symbolic reasoning. The key is to first construct a graph-based symbolic space that encodes the structure of the symbolic reasoning system, and then construct the parameter space of the machine learning model over this symbolic space. This approach will geometrically and algebraically lead to a space in which symbolic reasoning and machine learning are aligned. Since the target space and the model space are integrated to form a white-box system combining the characteristics of learning and reasoning, this should allow the results of machine learning to be explained by symbolic reasoning. Moreover, it is expected to achieve robust symbolic reasoning as an optimization in a continuous space.
JST Presto: Multisensory Integration in Biological Systems
Investigating the structure of multisensory systems using sensory-motor intervention,and applications of this in engineering
Even with recent advances in robot motion performance and AI, robots have not yet been successfully implemented in society. Unlike artificial systems, the behavior of living things in the natural world has a sense of flexibility and appropriateness. Most insects have only a few hundred thousand nerve cells in their whole bodies, making them clearly much smaller in scale than modern microcomputers, with their billions of transistors operating at speeds of tens of gigahertz. This leads us to believe that insects have an efficient system for functioning, completely different from artificial systems. If we could artificially recreate the structure of this insect system, we could develop an artificial system that is energy-efficient yet robust and adaptable. In other words, we hope to solve the problem of adaptability to unknown environments, which is a major reason why current robot systems have not gained ground in society, by artificially recreating the multisensory motor system of an insect. This study aims to (1) perform neuroethological experiments using a virtual reality (VR) system that can intervene in insect sensory and motor functions, (2) describe neural circuits using a computational neuroscience approach, and (3) develop motion algorithms that generate situation-adaptive movements from multi-sensory information using data-driven modeling, in order to thoroughly identify the multisensory systems of insects in the same way as mechanical systems. As well as identifying the adaptive behavior selection mechanisms of living creatures in response to multisensory stimuli, this will allow us to create an artificial system that can behave efficiently and adaptively in unknown environments with high uncertainty. This will contribute to understanding the structure of the incredible intellectual functions found in nature, and their applications in engineering.
JST Mirai Program
Engineerable AI Techniques for Practical Applications of High-Quality Machine Learning-based Systems
In this project, we are working on research and development of "Engineerable AI": technology to tailor AI systems to specific requirements, particularly in areas where safety and reliability are important, such as healthcare and autonomous driving. The focus is on technologies that can be utilized by engineers working on industrial applications of AI. There are two specific technical challenges in application of deep learning: the dependency on enormous amounts of data and the difficulty of aligning fine-grained prediction performance. First, AI depends on large amounts of data, making it difficult to build reliable AI systems for situations where data are scarce but important, such as diverse and rare types of cancer symptoms. Second, it is difficult to control the behavior of AI systems, which means that for real-world perception functions relating to safety, it is difficult to adjust performance for different situations while taking risks into account, or to make improvements while keeping the good aspects of previous behavior. In light of these technical challenges, in this project, we are working on techniques to embed domain knowledge in AI systems, and techniques to correct AI through analysis of AI errors. We will also provide a framework for comprehensively utilizing these technologies by thoroughly analyzing the needs and risks in the target domain and system. We are testing these Engineerable AI technologies in healthcare and autonomous driving, two examples of systems where safety and reliability are of vital importance.
JST START: Project promotion business support
Popularizing automated driving through technology providing logical explanations of software quality
This project aims to set up a venture company providing an ICT service as a business, to analyze and improve the quality of software and explain its safety to customers and society. Strategically focusing on automated driving in particular, we will provide technology to dispel safety concerns, enabling automated driving to be accepted and popularized within society. In contrast to the current approach of statistical safety assurance, there is a growing need for a logical approach providing strong safety assurance and high explainability. We will address this need by applying the advanced fundamental research results of the ERATO HASUO Project. Fundamental research into the technology to support the planned services and prototype implementation of the necessary software tools have already been completed during the main period of the ERATO HASUO Project (FY2016-FY2021). To develop these theoretical techniques as a business, research is required to improve the availability and portability of the technology, and the software tools need to be refined. This development will be carried out through this START project. As this is a highly advanced and unexplored technology, new research challenges may emerge in the course of business development. Thus, this START project will also involve research to resolve such issues. Through close collaboration with the ERATO HASUO Project (additional support period, FY2022-2024), which is carrying out fundamental and long-term theoretical research, we intend to accelerate both business development and academic research.
JST Presto: The Fundamental Technologies for Trustworthy AI
The Construction and Development of Risk-Aware Control Theory
Recent years have seen rapid progress in the automation of safety-critical dynamic systems that involve human lives, such as self-driving vehicles and aerospace systems. These systems must operate reliably even in highly uncertain environments. While numerous studies have focused on control technologies that enhance system reliability and safety by quantifying and incorporating uncertainty into design, existing control theories have failed to adequately address losses from rare, critical incidents in the design process. For safety-critical dynamic systems, however, it is essential to mathematically model both the probability and impact of unexpected, rare, and critical events. This modeling forms the basis for designing highly reliable control systems that balance performance and cost with risk tolerance.
In this research, we focus on developing a risk-aware control theory that addresses tail risks--events with extremely low probability of occurrence but potentially catastrophic consequences. This theory will mathematically analyze both the probability of occurrence and the impact of rare yet significant events. By incorporating these analytical results into control design, we aim to achieve levels of reliability and safety that were previously unattainable with conventional control theory. Furthermore, after establishing this new control theory foundation, we will reformulate key control problems, derive their solutions, and extend existing control methods into risk-aware control approaches.
JST FOREST
Spatial Control of Sound and Its Applications
Environmental noise has long been a social issue, but controlling the sound space is still an unresolved problem, as there are difficult technical challenges to overcome. High levels of noise in industrial environments, and even relatively low levels of noise such as road traffic noise and aircraft noise, are known to affect human health in various ways. Furthermore, issues with noise disrupting conversations and sleep in everyday life have become more noticeable now that more people work remotely on a daily basis. Many technologies aiming to reduce noise already exist. Active noise control, where noise is canceled by a loudspeaker driving signal, is known to be effective for low-frequency sounds that are often found in environmental noise. But the applications of this technique are limited to reducing noise in one-dimensional spaces such as inside a duct, or in very small areas such as noise-canceling earphones/headphones. The aim of this research project is to develop innovative technologies to reduce environmental noise by achieving active noise control in three-dimensional space, and generating sound zones, whereby sound is only played to people who want to hear it. Active noise control in three-dimensional space, or spatial active noise control, can only be achieved by simultaneously measuring and synthesizing sound spaces with high precision in real time. With an approach combining modeling based on wave theory with statistical signal processing, machine learning, and numerical optimization, we are researching the fundamental techniques of sound field analysis and synthesis to achieve spatial active noise control, and the applications of this technology.