Mariko Isogawa

Human behavior estimation using information that less lead to personal identification

Researcher

Mariko Isogawa

Outline

Towards a society in which people can be benefited from AI technologies without suffering any privacy issues, this project aims to build a human behavior estimation using information that less lead to personal identification. In particular, this project proposes a human behavior estimation that does not use information that leads to personal identification (e.g., images including users’ faces and clothes, users’ voices, or conversations). Instead, this project uses input that includes less personal information such as silhouette images, or signal waves without semantic information.

Kazumasa Uehara

Trustworthy Explainable AI understood by neuroinformatics technology

Researcher

Kazumasa Uehara

Outline

The goal of this project is to test the trustworthiness of Explainable AI (XAI) based on a causal approach via the human body. Using biological signals such as cortical neural activity and muscle activity obtained during human cognitive or motor task, AI (Deep Learning) first identifies the most important biological features that play a crucial role in the given cognitive or motor tasks. Using this feature extraction, I will then manipulate human cortical neural activity using non-inversive brain stimulation and manipulations of electromyogram signals using the cyber-space during the tasks. This project will further extend from laboratory to clinical research to build a prototype system of XAI, thereby contributing to the realization of XAI in the medical field.

Masako Kishida

Construction and Development of Risk-Aware Control Theory

Researcher

Masako Kishida

Outline

As more and more safety-critical dynamical systems become automatic, the reliability of control designs under various uncertainties is even more essential. However, existing control theories, while able to cope with uncertainties to some extent, are unable to adequately account for losses due to rare criticality accidents. This study aims to develop a risk-aware control theory that minimizes losses and ensures reliability and safety by quantifying tail risk and reflecting it in control design.

Ken Sakurada

Scene-privacy-aware Spatial Modeling

Researcher

Ken Sakurada

Outline

This project develops a unified spatial modeling method that does not depend on the type of sensors, such as a camera or laser, and an automatic editing function to protect the privacy and content of the scenes contained in the spatial model. These technologies enable the release and sharing of spatial models, which has been difficult until now, and encourage industries based on spatial data infrastructures, such as autonomous driving, service robots, and XR.

Yusuke Sakemi

Data-Driven Design of Neuromorphic Hardware

Researcher

Yusuke Sakemi

Outline

Neuromorchip hardware is a next-generation computer that achieves both high energy efficiency and high information processing capability. However, its operational reliability is low due to problems unique to analog circuits, such as non-ideal characteristics and variation characteristics of circuit elements. In this research project, by utilizing knowledge of the brain and data-driven modeling technologies, I aim to establish a design framework that integrates devices, circuits, and algorithms, and thereby construct neuromorphic hardware that can operate with high reliability.

Kazutoshi Tanaka

General manipulation primitives of autonomous soft manipulators

Researcher

Kazutoshi Tanaka

Outline

This research focuses on general manipulation primitives of autonomous soft manipulators. A soft body to try various motions safely, finding motions consistent with features of its body and an object, and an efficient method for generalizing manipulation primitives are needed for such manipulators. Thus, this research develops a soft, compact, and lightweight robotic arm, a method for generating motions representing features of an object, and an efficient method for generalizing the primitives.

Yoichi Tomioka

Realization of Sustainable Energy efficient AI Systems

Researcher

Yoichi Tomioka

Outline

In mission critical systems of infrastructure and medical systems, and so on, AI failures can cause serious malfunctions that can affect human lives, and thus AI must be fault tolerant. However, applying existing fault tolerant technologies to AI systems has problems that significantly increase area, power consumption, and cost. This project will establish a technology to realize a fault-tolerant AI that can detect sudden failures and continue to recognize them with sufficient accuracy with low computational power and a small area.

Futoshi Futami

Developing a learning theory for uncertainty in Bayesian deep learning using information theory

Researcher

Futoshi Futami

Outline

By combining information theory and PAC Bayesian theory, we aim to develop a new non-asymptotic learning theory that can simultaneously evaluate the prediction performance and uncertainty of Bayesian deep learning. Based on this developed theory, we will also present new Bayesian deep learning algorithms and models that can appropriately evaluate uncertainty and flexibly apply to users’ assumptions about the data.

Shingo Murata

Real-world robot learning based on free energy principle and play data

Researcher

Shingo Murata

Outline

We aim to establish the computational principle and data acquisition method necessary to realize robot learning in real-world environments with reference to the human cognitive development process. We will build a deep generative model based on the free-energy principle, which is considered a promising information theory of the brain, and implement it on a robot. By efficiently acquiring diverse robot’s play data, we will perform self-supervised learning and verify the robot’s ability to adapt to unexpected situations and plan for inexperienced goals.

Masaki Waga

Techniques toward reliable AI systems with quality assurance and explanation

Researcher

Masaki Waga

Outline

Toward reliable AI systems, it is necessary to guarantee and explain their “correctness,” such as safety, fairness, and robustness. This research combines automatic system approximation as a mathematical model and formal verification of its “correctness” to guarantee and explain the “correctness.” In addition to techniques for reliable AI systems, we aim to develop and publish prototypical tools.