Yusaku Abe

Sparse modeling for self-organization toward controlling pattern formation

Researcher

Yusaku Abe

Outline

This study aims on developing a method to control pattern formation of carbon nanotube self-organization process by using sparse modelling. The method developed would be able to link individual synthesis parameters and ultimately would describe the entire process. Therefore, this method is expected to lead the realization of a system that would comprehensively alter reactants rapidly and precisely to generate products with desired functions.

Kai Amino

The analysis of mimetic traits of butterflies via deep neural networks imitating predator

Researcher

Kai Amino

Outline

Mimicry is a fascinating phenomenon generated by natural selection and has attracted the interest of many people. However, measuring the perfection of mimicry is difficult. The conventional method is inevitably accompanied by arbitrary judgments of an observer and the assessment by real predators needs a costly experiment. In this project, I’m going to develop a novel deep neural network that imitates the visual system of birds and try to assess the quality of the mimetic trait of butterflies as an example. This project will contribute to our understanding about how can artificial intelligence imitate the reaction of the animal other than human.

Hwajin Lim

Research on urban management systems for the formation of urban collective knowledge driven by AI

Researcher

Hwajin Lim

Outline

This research suggests an urban management system for urban collective knowledge driven by machine learning-related AI(Artificial Intelligence) that can actively promote understanding the local needs of residents and participation in urban management. Specifically, this research proposes a model case for urban management by (1) collecting QoL(Quality of Life) data and integrating it with existing data by digitizing analog data, (2) building and operating a model of QoL index using AI, and (3) calculating dynamic indexes at multiple spatial and temporal scales such as district, community, and wide-area using Geographical information system (GIS).

Emiko Uchiyama

AI for Understanding Visuospatial Cognitive Functions in the Light of Spatio-Temporal Representation and Sensory Integration

Researcher

Emiko Uchiyama

Outline

The purpose of this research is to clarify the human cognitive process by constructive approach under the assumption that the human cognitive process is explained as a unified model constructed by several modules. This research especially focuses on the symptoms of functional disabilities of dementia patients. Using AI based on the model found out in this project, the following achievements are expected to gain; (1) clarifying effects on motion execution and cognitive function disabilities using the spatial and temporal representation in the model and (2) explaining these disabilities from the views of the dullness of their time perception and the declining of their self-agency due to changes in the width of the time window of the sensory integration.

Satoshi Ota

Systemized theory of magnetization dynamics of magnetic nanoparticles by pattern recognition

Researcher

Satoshi Ota

Outline

Magnetic nanoparticles have attracted attention as nanomaterials controlled in a non-contact manner by magnetization dynamics under a magnetic field. In this study, experimentally observed magnetization dynamics are assessed by a pattern recognition. The functions reflecting magnetization dynamics are prepared for development of magnetic nanoparticles by taking into account the interactions among parameters associated particle structures, surrounding environments, and applied magnetic fields. The theoretical system to design magnetic nanoparticles for future applications are constructed.

Ayumi Ohnishi

AI-based Sensation Enhancement to Overcome Problems Derived from Exhausted Five Senses

Researcher

Ayumi Ohnishi

Outline

Sensory enhancement devices, such as eyeglasses and hearing aids, are essential for improving the quality of human life. There is a possibility that the human five senses are daily affected by fatigue. If a sensory enhancer does not consider fatigue, it causes serious problems. This research investigates the effects of physical and mental fatigue on the five senses. Based on these results, I implement the AI-based sensation enhancement system to overcome problems derived from exhausted five senses.

Masateru Kawakubo

Diagnostic support AI system for myocardial ischemia by patient-by-patient basis blood flow information

Researcher

Masateru Kawakubo

Outline

Impaired coronary blood flow causes myocardial ischemia. Optimal treatment of myocardial ischemia is based on the image diagnosis by experienced radiologist. Therefore, a lack of experienced radiologists leads to less confident treatment and can undermine the consensus and conviction of patients in clinical situations. In this research project, I am planning to build the AI that proposes additional diagnostic image information to the conventional myocardial imaging for ischemia. Also I will reveal that this diagnostic support AI will increase the confidence of diagnoses of ischemia in general radiologist and cardiologist. My goal in this project is to build a diagnostic support AI system which allows patients to be treated of myocardial ischemia with peace of mind.

Tatsuro Kawamoto

Research on classification of graph-associated text data

Researcher

Tatsuro Kawamoto

Outline

We theoretically study the performance of a graph-base survey framework that is suitable for open-ended questions. We study how the order constraints and perturbative effects contributes to the inference of the collected datasets in the survey framework.

Eisuke Sato

Machine Leaning Assisted Organic Synthesis: Optimization of an Electrochemical Microflow Process

Researcher

Eisuke Sato

Outline

In this project, the machine learning assisted organic synthesis with an electrochemical microflow reactor will be developed. To input non-numeric parameters, such as the information about starting materials and reagents, these parameters would be converted to numeric parameters. Then, the model will be constructed to predict the reactivity against each condition. The use of chemical structures as the input for the model construction would allow the prediction of the different reactions. Especially, the construction of the model, which could predict the starting material dependent suitable reaction conditions, is the first target of this project.

Koya Sato

Exploring Wireless Resources via Image Sensor Based Crowdsensing

Researcher

Koya Sato

Outline

This research project aims to design a fast and accurate method for estimating radio propagation using deep learning and image sensor based crowdsensing. The communication efficiency of wireless communications is limited by radio propagation estimation accuracy. The research motivation is in the fact that radio propagation characteristics strongly depend on structure information. To break through the performance limitations of the conventional received signal power based crowdsensing, this project establishes a method for obtaining and using structure information specialized for estimating radio propagation from video and images. The biggest deliverable obtained through this project will be a data analysis infrastructure that can meet the growing demand for wireless communications over the next generation.

Yuhei Shimada

Designing Law and Policy for Telemedicine with Next Generation IoT and AI

Researcher

Yuhei Shimada

Outline

In order to realize the next generation of society utilizing IoT and AI, we need legal policies that support the development and implementation of the technologies. This study will demonstrate the innovation and systematization that is occurring in the field of telemedicine, and identifies the factors that create such an incubative environment both at the individual and regional levels. In addition, aiming at resolving conflicts between traditional legal systems and social change, and sublimating the identified factors into institutions, this study will construct policy packages.

Asuka Suzuki

Creation of high functional cooling components based on data driven structural optimization

Researcher

Asuka Suzuki

Outline

Addtitive manufacturing technologies permit fabricating more complex shaped cooling components than existing components. In the present study, I aim to optimize the structure of the cooling components by a data driven structural optimization. The model and property data obatined by thermal fluid finite element analysis of complex structures are used for training data of convolutional neural network. Using the neural network model, the optimized structure is explored based on the bayesian optimizaiton. Better cooling properties of the optimized structure are demonstrated experimentally.

Tsubasa Tanaka

Integration of rule-based methods with machine learning in computer-aided music composition

Researcher

Tsubasa Tanaka

Outline

In the field of computer-aided music composition, how to reflect composer’s intention, give originality to the generated pieces, and guarantee the pieces’ quality are important issues. To deal with these problems, composers tend to prefer rule-based approaches based on objective functions or constraints of their own design rather than machine-learning-based methods. With this in mind, this project aims at integrating rule-based approaches with machine learning to utilize the power of machine learning in composition systems.

Takao Dantsuji

Urban Transportation System Optimization via Traffic Flow Theory and Reinforcement Learning

Researcher

Takao Dantsuji

Outline

AI approaches utilizing a wide variety of traffic data are expected to be applied for the traffic congestion mitigation. However, storing a large amount of the row traffic data is difficult in terms of the storage capacity. Effective data management strategies are required. This study aims to develop a novel framework for integrating the effective data management based on traffic flow theory and transportation demand management optimization via reinforcement learning.

Jingfeng Zhang

Discouraging adversarial attacks through improving the adversarial training

Researcher

Jingfeng Zhang

Outline

Adversarial training (AT) is a trendy training style that can effectively defend against adversarial attacks. This research proposal will discourage adversarial attacks by improving AT through the three perspectives: a) network structure, b) defense prioritization, and c) attacker’s incentives.

Chia-Ming Chang

Design Thinking for Facilitating Data Annotation and Machine Learning

Researcher

Chia-Ming Chang

Outline

There is a huge demand for better machine learning results. One popular approach is to design good algorithms which many studies focus on. However, there is also another approach which is to “provide better data”. This study focuses on how to provide better data (without increasing costs) for improving machine learning results. This study aims (a) to bring “design thinking” into machine learning, (b) to improve machine learning results by providing better data, and (c) to design new annotation interfaces that allows non-expert annotators to complete a labeling task more quickly, more correctly, and support self-learning during annotation.

Mingbo Cai

Learning categories grounded in sensation without supervision

Researcher

Mingbo Cai

Outline

In most current deep learning works, categories are learned by supervised learning, and the resulting neural networks learn to categorize images instead of objects. To overcome this, this project will develop new neural network architecture and learning objective to force neural networks to emerge categorical representations of objects by learning to predict their future states without relying on human supervision. It takes inspiration from how infants and animals form initial concepts of categories by themselves.

Narumasa Tsutsumida

Multiscale and multiangle remote sensing data integration

Researcher

Narumasa Tsutsumida

Outline

There are diverse types of remote sensing data available , but they are not well organized. This project aims to integrate multiscale and multiangle remote sensing data to map land covers. With the use of aero photos and street-level photos, an approach that enables estimating land cover/land use information and building a geospatial database will be developed. Then, this ground information will be utilized to map land covers with satellite-based remote sensing data.

Yuri Nakao

Creation of New Philosophy of Technology Based on AI-Human Interaction

Researcher

Yuri Nakao

Outline

The goal of this project is to create a new philosophy of technology that takes into account the changes of the values in both humans and technology as the foundation for considering what technology should be. To this goal, the project representative will develop and evaluate an AI tool that changes accordingly to values in people. By developing the tool that ensures human autonomy using the framework of Responsible Research and Innovation (RRI), and describing and modeling the value changes based on tool usage logs, the project representative will analyze the values and norms in humans that change during the use of technology from the perspective of the philosophy of technology.

Kazuya Nishimura

Development of a biomedical-image recognition method that utilizes the ability of deep learning to learn relevant tasks

Researcher

Kazuya Nishimura

Outline

In the biomedical field, some labels can be easily collected with few amounts of effort that are called weak labels. The research aim is to develop a novel biomedical-image recognition method using such weak labels by effectively utilize the learning ability of deep learning. The research helps to reduce the annotation cost of AI techniques. And, it is expected to lead to further utilization of AI technology. I will start research with three goals: A. cell type classification of each cell; B. cell shape recognition for each cell; C. effect recognition of growth factor.

Chie Hieida

Research on emotional response models toward robots having emotions

Researcher

Chie Hieida

Outline

In this study, computational models of emotional responses are developed. More specifically, emotional responses are measured under two different conditions of visual and auditory stimuli, and computational models of emotional responses that take individual differences into account are examined. Although there are various definitions of emotional responses, this study refers to them as the physical responses such as visceral activities that occur in response to stimuli. This project aims to elucidate basic mechanisms behind emotions and to open up a path toward building robots with emotions by computationally modeling the emotional responses, which are the base of emotions.

Tatsuya Hiraoka

Exploring Tractable Tokenization for Both Human and AI

Researcher

Tatsuya Hiraoka

Outline

This project compares the unit of a word that is tractable for both humans and AI and analyzes their nature from the viewpoint of linguistics and machine learning to find common parts of them. The project also aims to develop a new tool for natural language processing to tokenize a sentence into units of words that are tractable for both humans and AI. The researcher is going to confirm that the developed tokenization tool contributes to the improvement of the performance on tasks of natural language processing as keeping the understandability of the units for humans.

Takashi Morita

Exploring deep transfer learning across heterogeneous data

Researcher

Takashi Morita

Outline

Today’s artificial intelligence (AI) is built on big data. When the primary data for AI training is not available on a large scale, a method called transfer learning is used; AI is first pretrained on other big data that look similar to the main data, and then, a part of the pretrained AI is fine-tuned for the main data. However, even such “similar” big data are often unavailable. To tackle this problem, this research project extends transfer learning to pairs of heterogeneous data, which do not look alike on surface, and explores its effectiveness. Heterogeneous transfer learning also provides a novel similarity metric between data whose comparison is of particular academic interest (e.g., human language vs. animal vocalization).

Yuya Morimoto

Machine-learning-driven ultrafast electron-beam control

Researcher

Yuya Morimoto

Outline

Electron beams have been widely used in both science and industry. Even though the spatial shape of electron beams can be precisely controlled, it is still challenging to control the temporal structure with high precision. In this research, we aim at developing a technology for producing electron beams having tailored temporal profiles with the aid of machine learning (AI). This research will open up novel opportunities for ultrafast imaging of chemical reactions, quantum information, damage-reduced electron microscopy and particle accelerators.

Hiroaki Yamada

Innovation in social simulation through fusion of machine learning and social science

Researcher

Hiroaki Yamada

Outline

Social simulation has potential to innovate business process and organizational decision-making in fields such as pedestrian flow management, traffic flow management, and logistics management, but there is a critical problem that modeling of them is extremely expensive. This project aims to establish an automatically modeling framework in the social simulation, through developing graph neural networks model and novel algorithms estimating social interaction structure.