*This Program finished in FY2021
　Active Period: FY2016 - FY2021
　Affiliations and titles are at the end of each research activity.

Ryo Akiyama

Controlling Object Appearance through Light Projection Induced Visual Illusion

Researcher

Ryo Akiyama

Outline

Human brain processes colors relative to their environment. As a result, they can be perceived differently from their physical components. For example, although a tree leaf reflects both green and orange light rays during dusk, we still perceive its color as green. In our work, we utilize this feature of human vision to control color of objects through light projection. Our aim is to employ this intelligent light equipment to further explore unknown areas of human visual system and better understand its characteristics.

Daichi Amagata

Fast Neuron Data Retrieval based on Spatial Data Modeling

Researcher

Daichi Amagata

Outline

Recently, many works have begun to figure out the mechanism of neurons and model them to simulate their actions and interactions on machines, which makes easier analysis of them. Because the cardinality of neurons is huge and their structures are complex, a technique to achieve fast simulation is required. Motivated by this, we deal with neuron as spatial data, and develop efficient techniques that mine some relationships between neurons.

Ari Hautasaari

Supporting Non-Native Speaker Contribution in Multilingual Collaboration

Researcher

Ari Hautasaari

Outline

This research aims to alleviate the problems non-native speakers of a language experience when trying to contribute their comments, opinions and ideas to the other members of a multilingual group in two modalities: face-to-face meetings and text-based group chat. The research approach focuses on exploring how to automatically detect non-native speakers’ status during meetings, share this information with native speakers, and influence the behavior of group members to include all participants equally in the collaboration. The objective in this project is to develop supporting technology for multilingual meetings where non-native speakers are a minority.

Miki Ueno

Developing a Dataset and Framework to Analyze Stories in Four-scene Comics Using Deep Neural Networks

Researcher

Miki Ueno

Outline

In this research, I focus on the multi-modal medium of four-scene comics, which depict a clear structure of stories, to represent states and causal relationships using natural language and pictures. I construct an original dataset and framework by integrating elemental technologies of machine learning to analyze four-scene comics. My aim is to share Japanese comic representation with different parts of the world to enhance communication and expand the possibility of co-creation with more people and computers.

Yoshitaka Ushiku

Creation of technology to automatically generate captions to various data

Researcher

Yoshitaka Ushiku

Outline

For image / video caption generation, which is one of the ultimate forms of media understanding, there are three requirements, (i) correspondence to individual users, (ii) correspondence to detailed expressions, (iii) data without ground truth captions. Although it is inherently necessary to deal with it, it is still unattainable. In this research, we integrate advanced knowledge in various fields such as natural language processing, image recognition and machine learning, realize these functions, and enable caption generation from various kinds of data.

Mayu Otani

Story understanding from videos and text

Researcher

Mayu Otani

Outline

Most effort in video understanding has been devoted to video captioning and video question answering. However, most of the problems provided by these tasks can be solved by understanding the content of one frame in a video. In this research, we will propose a new task focusing on stories expressed in videos as the next goal of advanced video understanding.

Yasushi Kawase

Algorithmic Research for Fair Allocation

Researcher

Yasushi Kawase

Outline

Allocation problem is to find an allocation of items to agents. The purpose of this research is to construct algorithms that find an (approximately) fair allocation efficiently. In particular, we explore the use of techniques from online algorithms to provide algorithms for allocation problems such as stable matching and combinatorial auction.

Shuhei Kurita

Knowledge extraction from large corpora and anaphora resolution through deep learning

Researcher

Shuhei Kurita

Outline

In regular conversations, we often omit entities that are implied when forming sentences. In order for computers to understand human conversations, they require anaphora resolution to infer omitted entities. In this research, we attempt to enhance the accuracy of anaphora resolution tasks with knowledge extraction and generation of pseudo training data from large unlabeled corpora, using deep learning techniques. We also propose methods to apply our techniques to machine translation and dialogue systems.

Atsushi Keyaki

Information retrieval system using semantic information

Researcher

Atsushi Keyaki

Outline

I aim at developing an enhanced information retrieval (IR) system, in which we extract semantic information based on a natural language processing technique and take each user’s latent information needs and search intention into account. More specifically, our work achieves a “controllable” machine learning-based IR system by comparing search accuracies between traditional and cutting-edge machine learning-based IR techniques.

Tsutomu Kobayashi

Software Microscope: A System for Consistency-Preserving Automatic Abstraction of Formal Specification

Researcher

Tsutomu Kobayashi

Outline

Formal specification methods have attracted strong interest because they enable human developers to not only systematically verify but also understand properties of target systems. However, in reality, understanding, modifying, and reusing formal specifications are difficult due to the large size and complexity of contemporary software systems. To address this problem, we develop a system to automatically construct abstract versions of given specifications (and proofs of their properties), which is consistent with the given specification and focuses only on particular elements of target systems.

Yusuke Kobayashi

Theoretical Development of Algorithms in Shrinking Networks

Researcher

Yusuke Kobayashi

Outline

There are a lot of theoretical and practical studies on optimization problems in networks that are models of communication networks and transportation networks. In this research project, we deal with networks that cannot maintain the current scale. For such shrinking networks, we introduce mathematical models of “good” networks and develop a theory of algorithms to obtain “good” networks.

Shigeyuki Sato

An auto-tunable framework of generalized N-body algorithms

Researcher

Shigeyuki Sato

Outline

Generalized N-body algorithms can work as a powerful programming tool for solving a wide range of application problems. Their potential applicability is, unfortunately, offset by considerable programming burden in implementing them efficiently for a specific problem on a specific parallel machine. To resolve this problem, we develop a domain-specific language for generalized N-body algorithms that is implicitly compiled to efficient parallel implementations and an auto-tunable runtime library.

Hiroaki Shiokawa

Time and Space Efficient Big Data Processing via Data Skewness Caching

Researcher

Hiroaki Shiokawa

Outline

Large-scale data analysis has been successful in a wide range of fields such as business, medical care and so on. However, a high-performance computer is indispensable so as to analyze such large-scale data quickly. The goal of this project is to establish time and space efficient algorithms that can handle large-scale data on the limited computation resources. The key idea of this project is to design data processing algorithms as to reduce redundant computations by using frequent data structures hidden in the data. With our proposal, we aim to improve feasibility of the large-scale data analytics.

Ryo Suzuki

The development of a visual programming language for making information technology accessible to everyone.

Researcher

Ryo Suzuki

Outline

This project aims to develop a new visual programming language “Enrect” that assists everyone to access information technology in their own everyday life regardless of age, mother tongue, and physical ability. By combining a design that improves the understandability of the code, an interface best suited for touch operation, and a rich feature set, development of practical applications will be achieved with less time and learning cost than before.

Tasuku Soma

Efficient and unified algorithms for online submodular optimization

Researcher

Tasuku Soma

Outline

Recently, submodular optimization has attracted interests in machine leaning and network science. However, traditional submodular optimization cannot deal with uncertainty in optimization problems. In this project, we will apply techniques in online optimization to submodular optimization and establish a new framework “submodular optimization under uncertainty”.

Tsuyoshi Takatani

Rendering Translucency and Glossiness on Digital Fabrication

Researcher

Tsuyoshi Takatani

Outline

Digital fabrication is developed in both industrial and academic fields and utilized for various applications. However, main research topics are for shape and functionality but not for appearance. This work focuses on the translucency and glossiness of a fabricated object to render richer appearance as the purpose of this work. The key idea is combining 3D printer and UV printer to design the behavior of light on and inside of the object.

Chenhui Chu

Image-pivoted paraphrase extraction for deeper natural language and image understanding

Researcher

Chenhui Chu

Outline

A paraphrase is a restatement of the meaning of a text in other words. In this work, we extract paraphrases from images and their multiple related text using image regions as a pivot. It especially enables the precise extraction for paraphrases of the same concrete concept thanks to the mediation by images. The extracted paraphrases will significantly improve the performance for not only natural language tasks such as information retrieval but also natural language and image multimodal tasks such as visual question answering.

Ran Dong

Interaction between humans and puppets: Emotional motion expression design Using Japanese Bunraku puppets

Researcher

Ran Dong

Outline

Nowadays, Artificial intelligence assistants such as Apple “Siri” cannot become popular in human life. The reason can be considered as we can not feel the human emotion from those assistants. This research develops a communication robot using Bunraku puppets’ sophisticated emotional motion expression. We analyze difficult human emotional motion expression to makes communication home robot can express emotional motion artificially.

Toru Nakashika

Vocoder-free Non-parallel Voice Conversion Using Complex-valued Adaptive Restricted Boltzmann Machine

Researcher

Toru Nakashika

Outline

Voice conversion (VC) is a technique that enables articulation disorders and laryngectomized patients to enounce with their voices. There are two mainstreams in the VC technique: parallel and non-parallel approaches. Both approaches have their own issues: the former requires large amount of speech data, and the latter tends to degrade the converted speech due to use of a vocoder. The goal of this study is to propose a non-parallel VC method without using any vocoders, which overcome the two issues. I seek the best way to put VC into practical use and to help the people with oral disorders.

Takayuki Nishio

Learning Based Self-optimizing Wireless Networks

Researcher

Takayuki Nishio

Outline

This research aims to develop self-optimizing wireless networks, which optimize control parameters adaptively and maximize network performances so as to provide stress-free wireless access. The research studies a system architecture and learning scheme to acquire optimal wireless control.

Ryohei Hisano

Property Graph Approach to Incorporating Commonsense in Social Data Analysis

Researcher

Ryohei Hisano

Outline

By storing various social data in a unified graph database and developing a graph algorithm over it, we aim to practice cross data analysis. From our approach we could, for instance, analyze inter firm buyer-seller network also taking into account various other hidden relationships behind the network. These hidden relationships includes what goods each firms produce, their meronyms, social network among CEOs, etc. which presumably carries important information over how we interpret the inter firm buyer-seller network. In the future, we aim to develop methods to reason over the database so that we could incorporate human level common sense.

Shuichi Hirahara

Study on Restricted Circuit Minimization Problems and Circuit Lower Bounds

Researcher

Shuichi Hirahara

Outline

A computer consists of logic circuits, which are built by composing small components such as AND and OR gates. In order to build small hardware, it is important to minimize the number of gates. This task can be formulated as mathematical problem known as “Minimum Circuit Size Problem”, and it is one of the most central problem in computational complexity. This study aims at investigating the computational hardness of the problem.

Yosuke Fujii

Data-driven shape and motion analysis of two-photon microscopy image data

Researcher

Yosuke Fujii

Outline

Two-photon microscope helps us observe live images of cell shape and motion. However, analyzing such imaging data is difficult. Imaging researchers make ad hoc features to analyze imaging data. In this research, data-driven feature extraction methodology of shape and motion will be developed from the standpoint of mathematics, statistics, and computer science. Feature extraction and quantification of shape and motion of cells makes it possible to integrate imaging data and omics research.

Matthew James Holland

Safe AI is efficient AI: improved generalization via robust learning algorithms

Researcher

Matthew James Holland

Outline

If machine learning technology is to become a central part of modern social infrastructure, it must be sufficiently reliable “in the wild,” where a high degree of autonomy, economy, and precision is required. Satisfying these conditions simultaneously is a challenging learning task, and modern learning algorithms look to be reaching their limits. To break past existing limitations, we develop a new class of learning procedures, characterized by an adaptive feedback mechanism, which prioritizes robustness and stability. By paying the small price of some minor computational overhead for improved statistical estimation, we can realize major savings on net learning time, all while dramatically reducing the variance of the learner’s output.

Haruka Matsukura

Inverse Analysis of Chemical Signal Flow By Machine Learning

Researcher

Haruka Matsukura

Outline

The aim of this project is to establish a method that enables generation of a high-resolution gas flux map in a field of several tens of meters square from a limited number of gas concentration measurements. Gas is distributed in the field, following the diffusion-convection equations which show strong nonlinearity. It is difficult to perform the inverse analysis of the gas transport to estimate gas emission sites and the amount of gas flux. To overcome the strong nonlinearity problem, machine learning techniques are introduced in this project. The potential applications include monitoring of landfill gas emission and search for landmines.

Takashi Miyamoto

Detection of earthquake damage of housing structures from SAR satellite images using convolutional neural network

Researcher

Takashi Miyamoto

Outline

This study aims to achieve high accuracy in extraction of housing damage from SAR satellite images by applying approaches of characteristic extraction and discrimination based on a deep learning model to image analysis obtained immediately after earthquake. Upon identification of positions of housing structures within the images achieved by an integrated processing of satellite images and geospatial information, a damage discrimination with high accuracy is achieved by applying convolution neural network to pixel groups including houses.

Kazuya Murao

Annotation Method for Sensor Data using User’s Unresponsiveness to System Notifications

Researcher

Kazuya Murao

Outline

Realizing high performance human activity recognition (HAR) technology using smartphone and wearable device requires large amount and wide variety of sensor data labeled with groundtruth, which expresses the situation the sensor data was captured. This research tackles on the construction of annotation method that ties data and groundtruth from how the users deal with notifications that occurred on the device. The proposed method contributes to the realization of the environment where groundtruth can be captured along with sensor data, resulting in the progress and prevalence of HAR technology in the future.

Shogo Yamashita

Developing a 3D Water Flow Measurement Technology for Swimming Pools

Researcher

Shogo Yamashita

Outline

The objective of this research is to develop a novel three-dimensional (3D) water flow measurement technology for swimming pools. This technique increases training efficiency in water sports and produces interactive video content in projection-based entertainment swimming pools. The technology has the potential to achieve a wider measurement field and has low adverse effects on the human body and sight, all of which are required for water flow measurement in swimming pools.

Hironori Yoshida

Participatory Architectural Design and Fabrication with Irregularly Shaped Waste

Researcher

Hironori Yoshida

Outline

Diverse natural materials such as stones and woods have been used as architectural elements preserving their native forms since primitive shelters, however, the use of them in modern buildings is limited due to their irregular properties. In this project, we take the diversity as playful inputs for design task, and develop a human-in-the-loop workflow for design and fabrication of architectural elements. Taking tree branches as a material in native forms, we aim to fabricate networks of branches and realize it by CNC milling process. Each connection has a customized unique joint adapted to the native forms of branches. Together with low-cost mobile scanning devices and collaborative interface between users and digital fabrication tools, users can easily contribute to the process of making, which was an advantage of traditional architecture with local materials.

Lijun Liu

Development of a method for high-performance multiscale analysis with time-space parallel computing

Researcher

Lijun Liu

Outline

In this research, I propose to develop a fast multiscale computation method enable to coupling a microscale molecular dynamics model with a macroscale finite element analysis model. To realize a high-speed multiscale computation, I introduce the concept of hybrid time-space calculation which is challenging. In the parallel computing, not only the conventional space decomposition approach but also the state-of-the-art parallel in time algorithm are applied. The proposed method will be useful in various fields as a material analysis tool.