[AI powered Research Innovation / Creation]Year Started : 2020

Affiliation and job title should automatically appear from the information that a researcher registered with researchmap.
Data may be outdated or undocumented.
When there is not a connection via the internet, data are not displayed.

Kazuya Ishitsuka

Analysis of the distribution and properties of the Earth’s resources using AI

Researcher
Outline

Uncertainty in the subsurface conditions and properties is a risk for the development of underground resources. For the sustainable development of Earth’s resources, this study aims to develop a method for estimating the distribution of subsurface conditions and physical properties such as temperature and permeability and a method for elucidating the mechanism of rock properties using AI. By incorporating the laws of physics into the AI, the method will provide more reliable predictions than existing methods and contribute to the understanding of physical properties of rock materials.

Kazumasa Uehara

Decoding neural information on interpersonal coordination in skilled motor behavior using brain hyperscanning and deep learning-based analysis

Researcher
Outline

Interpersonal coordination appears in many situations of everyday life and is essential for our civilization. However, its neural mechanisms have a complex architecture and are still debatable. By combining a brain hyperscaning paradigm with a deep learning-based data analysis, this study works on decoding neural information underlying interpersonal coordination in skilled motor behavior from experts who are talented athletes and musicians. This result will contribute to a better understanding of neural mechanisms underlying interpersonal coordination in skilled motor behavior. Furthermore, based on the research findings, I attempt to build a neurofeedback system using experts’ neural information that enables unskilled individuals to improve interpersonal coordination. Through this project, I will propose an efficient and user-friendly smart learning system for interpersonal coordination.

Masahito Ohue

AI-based peptide design for controlling proteins

Researcher
Outline

The drug R&D process is becoming more and more complex every year, leaving only challenging drug targets. In this study, I focus on protein-protein interactions (PPIs) as the most challenging targets and develop AI methods to design peptides that can inhibit PPIs. I mainly aim to develop the techniques that utilize information on the PPI interface structure, which has not received much attention so far.

Tarin Clanuwat

On-Device Kuzushiji Recognition App for Historical Document Investigation

Researcher
Outline

Japan had been using Kuzushiji for over a thousand years. However, comparing to the large number of documents survived, the number of people who can read Kuzushiji properly is only a few thousands of the whole Japanese population. In order to tackle this problem, the project representative will examine on-device machine learning algorithm and develop Kuzushiji recognition smartphone app. This app will aid anyone to easily investigate their own materials. At the end of the research, this app will be released as a free app on both Google Play and App Store.

Kento Kawaharazuka

Innovation of Growing Robots based on Learning Theory of Informatized Body Model

Researcher
Outline

We will develop a learning theory of informatized body models that express the causal and spatial relationships among the body sensors in a robot, and its application techniques to innovate and create new growing robots whose internal systems and body structures are formed and developed from experience. Without human direct description, the robot self-generates controllers and recognizers from experience, updates and adapts them successively, deletes and adds body sensors appropriately, and restrains and expands the body using tools and the environment.

Jun Kitazono

Extension of integrated information theory based on submodularity and its applications to neuroscience

Researcher
Outline

Integrated information theory, a hypothesis about consciousness, states that our consciousness arises in a ‘core’ of the neural network, where integration of information between neurons is strongest. In this study, we generalize the concept of core based on a mathematical property called submodularity. By analyzing neural data with the generalized core, we attempt to understand where consciousness arises in the brain. We also apply the generalized core to a wide range of neural network analysis.

Akira Kusaba

Process design innovation in next generation semiconductor development

Researcher
Outline

To shorten the development period of next-generation semiconductors, it is necessary to quantitatively predict the crystalline quality from process conditions. However, the current multi-scale simulators composed of physical models in the gas, surface, and solid phases might predict the number of orders of magnitude or the direction of change in the target variables. In this study, I will improve the physical models with data from the state-of-the-art measurements and further integrate each physical model in a data-driven manner to achieve quantitative predictions and innovate process design.

Kyo Kutsuzawa

Imitation Learning That Results in High Generalization Ability Using Shared Synergies

Researcher
Outline

Imitation learning, in which robots learn human behaviors, is expected to be useful for automating human’s daily tasks, but it is necessary to reacquire human movements to adapt to new situations. This project aims to develop an imitation learning system that utilizes muscle synergies that are considered the neural basis of human movement generation. Using the knowledge that humans reuse common muscle synergies to generate movements even when the environment changes, this project aims to realize a system that reproduces humans’ generalization ability.

Nahoko Kuroki

Quantum chemistry on the chemical reaction fields: The development of artificial intelligence for precisely detecting chemical time-space evolution

Researcher
Outline

The development of novel chemical reactions with low environmental loads is indispensable toward a sustainable society shortly. To design chemical reactions, it is necessary to track not only “the electronic state of the reactants (the solutes)” but also “the reorientation dynamics of the surrounding solvent molecules during the reaction”, of which time scales are quite different. In the ACT-X project, I will develop a novel physicochemical technique to understand the time-space evolution of the chemical reaction fields, a hidden leading actor of designing chemical reactions, by making full use of physicochemical theory simulations and information science techniques.

Hiroyuki Kuromiya

Development of Evidence Ecosystem in Education

Researcher
Outline

The purpose of this proposal is to build a learning analysis based platform “evidence ecosystem” that systematically supports creating, communicating, and using evidence by utilizing educational big data. This system aims to automatically extract, classify and recommend evidence from learning logs just in time by using artificial intelligence techniques. In addition, from the perspective of educational practice, we will also identify the needs for evidence-based practice in the study field. It encourges teachers to use evidence for their decision making in their daily teaching activities.

Kojima Yasuhiro

Estimation of cell state dynamics based on the convection-iffusion process

Researcher
Outline

In this study, we describe the environment-dependent dynamics of the cell state by the Kolmogorov equation to construct a stochastic model linking the spatial transcriptome and cell dynamics, and develop a new algorithm for efficient optimization. This will provide a data-driven way to capture the average variability and fluctuations in the environment-dependent cellular state and estimate its molecular mechanism. In addition, we will elucidate the contribution of microscopic cellular state transitions to the fate determination of entire biological tissues.

Kayoko Kobayashi

Physical property prediction of wood by comprehensive analysis of multi-scale structure

Researcher
Outline

This project works on predicting wood physical properties with an artificial intelligence model to demonstrate its usefulness in wood science. Since wood has a complex and multi-scale structure, the structure-property relationships has not been fully understood yet. In this project, therefore, the data of wood structure is collected from multiple scales, from molecules to tissues, and the prediction models of the physical properties are constructed with machine learning. In addition, the structural features that determines the physical properties are investigated by the analyses of the prediction models.

Yuji Saito

Data-Driven Sparse Sensing for Aerospace Development

Researcher
Outline

In the applications of aerospace engineering such as launch vehicles and satellites, optimal sparse sensor placement is an important subject in the performance prediction, the control, and the fault diagnostics etc., because of the limitations of installation, cost, and downlink capacity for transferring measurement data. In this project, the reliability of aerospace crafts by developing data-driven sparse sensing and deploying it will be improved, and the high-speed reconstruction of complex fields and anomaly pre-detection of its failure events will be achieved.

Yuki Shimizu

Shape Optimization of Permanent Magnet Synchronous Motors Using Machine Learning

Researcher
Outline

Electric motors, which convert electrical energy into mechanical energy, are used to power various products that run on electricity. There is a strong nonlinearity between the shape and performance of these motors, and the design of these motors requires repeated experiments and analysis, resulting in a long development period. In this project, therefore, I aim to reduce the optimal design time of motors by combining the technology of electromagnetic field analysis and deep generative models.

Takashi Tanaka

Using explainable AI to establish a soil crop system model

Researcher
Outline

This project fristly aims to qunatify on-farm spatiotemporal variability in soil properties, which is further treated as input data to establish a model for crop yield predicition using deep learning techniques. This project secondly aims to perform a sensitivity analysis within the expections to enhance the explainability of the deep learning model. Based on the new knowledge coming from the explainable deep learning model, empirical soil crop system model will be estalished to adapt conditions specific to each region.

Hikari Tanaka

Analysis of cell fate decision leading to YAP-dependent necrosis in neurodegenerative diseases

Researcher
Outline

In neurodegenerative diseases, neuronal cell death, an important pathological process, is a critical point resulting in the disease progression from a treatable stage to an untreatable one. Hence, it is vital to reveal some indication of cell death and prevent it. This study focuses on the prediction of the fate decision of YAP-dependent necrosis by AI technology. It will reveal the human-invisible intracellular dynamics in a single cell prior to YAP-dependent necrosis.

Sho Tsuji

Developing datasets of infant behavior that are exploitable by AI

Researcher
Outline

Human infants learn about the world with unrivaled speed and efficiency. A key challenge to implementing insights from infant development into research towards next-level AI is to measure infants’ behavior in a way that is exploitable by computational approaches. Central bottlenecks are the amount of data we can collect from infant participants and their automated analysis.
The present project will build an infrastructure towards collecting and automatically analyzing standardized and large datasets of infant gaze behavior. Our solution has three key features: (1) Moving infant data collection to remote, online methods; (2) Setting up a multi-lab international collaboration to ensure large-scale standardized data from a diverse population of infants; (3) Developing gaze coding algorithms for automatic data analysis.

Kouta Nakayama

Development of FG-NER/EL System in 31 Languages

Researcher
Outline

The construction of the dataset used for training an FG-NER/EL system is quite costly. Therefore, we construct the dataset by exploiting the link structure of Wikipedia. However, the dataset contains many false-negative labels and has become a bottleneck for developing the system. In this work, we focus on true-negative labels and mitigate the effect of the false-negative labels by learning the model from small amounts of reliable negative and positive labels to build a high-quality FG-NER/EL system.

Tatsuhito Hasegawa

Development of versatile fish catch detection platform towards big data in the fisheries industry

Researcher
Outline

In this study, we develop a large-scale fishery image dataset, hardware of an AI monitoring device that can capture and detect the fish catch, and software for accurate recognition of fish species, length, and number of fishes that works robustly in various fishing grounds. This research aims to construct big data of detailed fish catch information for the improvement of fisheries operations and innovation in fisheries research.

PARINYA PUNPONGSANON

Computational Food Textures for Enhancing Eating Experience using Food 3D printer and AI

Researcher
Outline

This project will tackle the challenges that improve the eating experience by integrating the food 3D printing technique with AI. First, this project will explore the relationship between food texture and multi-sensory perception and its relation to computational parameters to derive computational parameters. The result from the exploration will then translate the multi-sensory of food texture into a digital form, which allows digital fabrication to create an alternative food with a similar experience to the target food that the user wants to eat. This project would enable people with an eating disorder or allergic to experience the food they want to eat.

Motoyuki Murashima

New Challenge to Develop the Triboforcast Technology utilizing Artificial Intelligence

Researcher
Outline

In the present research, we take on the challenge of developing a new tribotester for the purpose of utilizing artificial intelligence. To this purpose, the tribotester has the unique characteristics of continuous optical-imaging, high-resolution measurements, and high-precision synchronization. The measured optical photographs and friction coefficient are used to develop neural networks that predict the friction coefficient on the imaging surface. Subsequently, we use a neural-network evaluation technology (e.g., Grad-CAM) to reveal important optical characteristics of the surface. This technology establishes a new friction model that incorporates emergence and hysteresis, which is difficult to build with previous research approaches.

Hiroaki Yamada

Explainable outcome prediction model for civil dispute resolution in the Japanese law system

Researcher
Outline

We develop an explainable outcome prediction model for dispute resolution under the Japanese civil/commercial laws by natural language processing technology. The model takes a target case description written in natural language and predicts the outcome and generates its explanation, including supporting reasons and grounds. The model is jointly trained for both outcome prediction and explanation generation tasks where the tasks share attended features. In the project, we formalize the outcome prediction and explanation generation tasks for the Japanese law system. We also build a large scale data set in which each case is annotated with its outcome and explanation. The fruits of this project should make procedures faster and more efficient in every civil dispute scene.

Soh Yoshida

Realization of a Platform for Various Values Retrieval

Researcher
Outline

This project creates a search technology that provides accurate information and the whole picture of social media containing various values in an interpretable way. A method to generate rankings that do not contain false information and a search diversification method to arrange information from various viewpoints into a single page of rankings are developed. Furthermore, by extending the search results with related information other than keyword matching, visualization of the search’s bias in a heuristic way is aimed.

Quick Access

Program

  • CREST
  • PRESTO
  • ACT-I
  • ERATO
  • ACT-X
  • ACCEL
  • ALCA
  • RISTEX
  • AIP Network Lab
  • JST ProjectDB
  • Global Activities
  • Diversity
  • SDGs
  • OSpolicy
  • Yuugu
  • Questions