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

Soramichi Akiyama

Transparent Acceleration of Application Execution using Next-generation Memory Devices

Researcher

Soramichi Akiyama

Outline

This research aims to automatically accelerate application execution by exploiting next-generation memory devices that provide faster accesses than conventional devices in return of errors (bit-flips) injected into retained data. By developing system software that controls the next-generation memory devices to fit spatial and temporal locality of applications’ error-robustness, we gain the maximum speedup of target applications with guaranteed error-rate to the final results of the applications transparently from the applications.

Shigemi Ishida

Development of Vehicle, Bike, and Pedestrian Detection Method Using Wireless Communication

Researcher

Shigemi Ishida

Outline

Vehicle, bike, and pedestrian detection is one of the core tasks in the ITS (intelligent transport system). This research aims to develop a vehicle, bike, pedestrian detection system using off-the-shelf wireless communication modules without any modification. Wireless communication suffers from environmental changes. We analyze the changes in wireless communication status to detect vehicles, bikes, and pedestrians on a road.

Masakazu Ishihata

An Unified Approach to Combinatorial Optimization Using Decision Diagrams

Researcher

Masakazu Ishihata

Outline

A combinatorial optimization problem (COP) is to find a subset of given items that satisfies a given constraint and optimizes a given objective function. COPs appear not only in real world problems but also as subproblems of artificial intelligence and machine learning. In general, COPs with different constraints and objective functions are solved by different algorithms. In this research, we aim to propose a unified approach for COPs based on decision diagrams (DDs). A DD is a compact graph structure representing constraints, and supports many queries such as linear function optimization, sampling, and the rest. We first attempt to expand the class of constraints that can be efficiently represented by DDs, and also expand the class of queries that can be efficiently executed on DDs. Then, we construct a unified method for COPs by using DDs.

Kenta Itakura

Development of progressive technique of plant measurement using machine learning and 3D image processing

Researcher

Kenta Itakura

Outline

Effective solution for air pollution and global warming is crucial for our future. It is vital to extract the function of their purification of plants with information technology such as machine learning. Automatic detection of plants from 3D images and synthetic analysis of plant physiological and structural information will be conducted. We realize a fully automatic system that extracts the function of plant environment purification. This method can be connected with automatic driving and drone, so that autonomous environmental protection can be done as a part of ICTenvironment.

Shinji Ito

Online Convex Optimization based on Partial Feedback

Researcher

Shinji Ito

Outline

Online convex optimization is a useful framework for sequential decision-making. This framework is, however, not applicable for the real-world problems in which only partial feedback is available. In this project, we consider online convex optimization problem with partial feedback. For this problem, we aim to develop algorithms that are optimal in the sense of the regret minimization, and aim to prove their effectiveness.

Danilo Vargas

Automatic Search for Robust Hybrid Deep Learning Models

Researcher

Danilo Vargas

Outline

In my previous research, it was shown that only one pixel is necessary to fool neural networks. This new finding show now only one vulnerability of neural networks but also that they are capable to understand images as we do. The origin of such vulnerability is in the model of neural networks. However, to find the most appropriate model and parameter takes time. Moreover, it is possible that deep learning models are not able to solve the current vulnerability. Consequently, this research will use an optimization algorithm to search automatically for robust hybrid deep learning models to create the next generation of deep learning systems.

Kodai Ueyoshi

Exploration of Hardware-oriented Deep Neural Network

Researcher

Kodai Ueyoshi

Outline

Energy saving deep neural network (DNN) processors are crucial to realize DNN processing under limited constraints. This research aims to explore efficient DNN models for the hardware structure. In particular, I focus on reducing external memory accesses which consume a lot of energy. It is expected that optimization of both energy efficiency and recognition accuracy by the cooperative design of algorithms and the hardware.

Yo Ehara

An Efficient Vocabulary Learning Support System that Considers Regions in Semantic Spaces

Researcher

Yo Ehara

Outline

Acquiring a second language is different from other types of learning in that learners can learn other specialties using the languages that they learned. Mastering specialties in a second language requires detailed knowledge of the language, for example, how to employ the newly learned vocabulary in diverse contexts. This study aims to develop a system that supports learners to efficiently acquire such knowledge by recommending and visualizing words to be learned and their usage. To this end, we construct methods to obtain distributed representations whose semantic spaces encode the semantic diversity of words’ contexts as regions.

Taihei Oki

An interaction of weighted combinatorial optimization and polynomial matrix theory

Researcher

Taihei Oki

Outline

A class of combinatorial optimization problems can be solved efficiently via matrix theory. Conversely, some problems on matrices are known to be solvable by using combinatorial optimization algorithms. In this project, we extend these established methods to a correspondence between weighted combinatorial optimization problems and polynomial matrix problems. In particular, we aim to elucidate how polynomial matrix theory can help to solve weighted combinatorial problems, and develop a new application of combinatorial optimization algorithms to polynomial matrix problems.

Masaru Oya

Integrated Circuit Design Specification against Hardware Trojans

Researcher

Masaru Oya

Outline

Modern digital integrated circuits (ICs) are often designed and fabricated by third parties/tools, which can make IC design/fabrication vulnerable to malicious modifications. The malicious circuits are generally referred to as hardware Trojans (HTs) and they are considered to be a serious security concern. This research develops secure integrated circuit design specification against HTs. The secure design specification can be verifiable by formal verification and detect HTs.

Tomoyuki Kajiwara

Personalized Text Generation based on Vocabulary Restriction

Researcher

Tomoyuki Kajiwara

Outline

Personalized text generation such as text simplification according to the reading skills of users and style transfer between formal and casual text, is crucial to facilitate communication between humans and machines. In this project, we formulate text generation as a task of representing the meaning of input sentences using a given vocabulary. We automatically restrict the basic and terminology vocabularies in order to generate personalized text.

Marie Katsurai

Development of a Method for Aggregating Research Accomplishments of Individual Researchers

Researcher

Marie Katsurai

Outline

Profiling research accomplishments (e.g. papers, research grants, awards, and dissertations) of individual researchers is necessary for evaluating their research activity and academic networking. Such accomplishments are generally scattered across separate databases, which have different schema. This study explores an approach towards author matching across different academic databases, automatically integrating the records and regarding them as an individual researcher’s accomplishments.

Yuya Sasaki

Research on Binary Decision Diagram for Large-scale graphs

Researcher

Yuya Sasaki

Outline

It is important for information science to efficiently handle discrete structures. Binary decision diagrams (BDD) enable efficient enumeration of all answers on discrete structure problems. However, it requires a large memory space, and thus it is hard to be applied to large-scale graphs. In this research, we develop a new BDD that can effectively compute an approximate answer set. In addition, using the BDD, we develop an efficient algorithm for computing the network reliability. Since the enumeration problems on graphs and the network reliability problem have wide applications, our techniques help to solve many problems in several fields.

Yuta Suzuki

Development of Data Analysis Method for On-the-fly Crystal System Prediction

Researcher

Yuta Suzuki

Outline

Efficient automated analysis of measured data is a critical issue in material development. In this research, we propose a new methodology that human cooperates with AI for automatic and accurate estimation of the crystal structure from X-ray diffraction data with the combination of X-ray diffraction simulation from materials database and machine learning. We will extract the identification rules that expert materials researchers used as tacit knowledge by analyzing the constructed machine learning models.

Ryo Suzuki

Dynamic Physical Interfaces

Researcher

Ryo Suzuki

Outline

This project explores Dynamic Physical Interfaces for the next generation of human-computer interaction. More specifically, I will explore the following three areas: 1) Dynamic 3D Printing for an instant and reconstructable shape formation, 2) Portable Shape Display for a thin and flexible dynamic tangible display, and 3) Programmable Architecture for a room that can dynamically change its shape based on human needs. Given these technologies, I will explore possible application scenarios in accessibility, personal fabrication, haptics for VR/AR, and education for the future of man-machine interfaces.

Koh Takeuchi

Spatio-Temporal Super Resolution from Multiple Low-resolution Spatio-Temporal Data

Researcher

Koh Takeuchi

Outline

Spatio-temporal data in real-world tend to be sparse and observed with a limited resolution due to the cost and the physical constraints. Such low-resolution data are not useful enough to analyze the dynamics of objectives and make predictions. To overcome this limitation, we develop a machine learning approach, Spatio-Temporal Super Resolution, which enables to estimate high-resolution spatio-temporal data from multiple low-resolution spatio-temporal data. Our end-to-end approach considers complicated relationships among multiple data with different spatial shapes and sampling rates and leverages them to estimate high resolution observations.

Teruo Tanimoto

New design method for flexible and power-efficient application specific processors

Researcher

Teruo Tanimoto

Outline

Application-specific processors are essential to meet area and power requirements in the post-Moore era. General purpose and application specific hybrid processor, which combines a general purpose processor and application-specific circuit, enables high power efficiency and flexible implementation under an area restriction. To reduce development cost, automatically detecting operations suitable to be processed in specific circuits according to deep analysis of application characteristics will be developed.

Han Bao

Development of Low Variance Risk Estimator for Weakly-supervised Data

Researcher

Han Bao

Outline

Recently, weakly-supervised learning, where incompletely labeled data can be used in the classifier training, has got attention in the field of machine learning, due to cost of data collection and privacy issues. One of the main approaches is to minimize an unbiased risk estimator. However, the variance of the risk estimator tends to be large and the learning process becomes unstable. This proposal is aimed at making properties of the risk estimator clearer to improve their performances, which is applied for real-world applications.

Tadanori Teruya

Fast Lattice Basis Reduction suitable for Massive Parallelization

Researcher

Tadanori Teruya

Outline

Lattice-based cryptography plays a major role in post-quantum and innovative cryptography. The security of cryptographic lattice-based schemes is based on the hardness of lattice problems. To assess security appropriately, research and development of efficient solvers of lattice problems are needed. Especially, lattice basis reduction is an important building block in the construction of faster solvers. The goal of this research is to develop fast lattice basis reduction algorithms suitable for massive parallelization.

Yugo Nakamura

Flexible context recognition mechanism self-adapting to human and environment

Researcher

Yugo Nakamura

Outline

Existing context recognition technologies which understand the human behavior and environmental conditions using various IoT devices are optimized for target environments. However, these existing technologies are suffering from recognition performance degradation because they can not adapt to daily life situations where the target users and the environment configuration constantly change. To address this problem, we aim to realize the flexible context recognition mechanism which self-adapts to changes of target users and environment configuration occurring in daily life by utilizing the accumulating knowledge.

Yu Nakayama

Adaptive Reconstruction of Telecommunication Network

Researcher

Yu Nakayama

Outline

Spatio-temporal patterns of human mobility result in significant fluctuations of mobile traffic. Such fluctuations drastically deteriorate the efficiency and financial viability of conventional mobile networks. To address this issue, this research investigates an adaptive mobile network architecture with moving nodes towards beyond 5G era. This work also targets network architectures and utilization methods based on crowdsourcing and sharing economy to further improve the adaptivity.

Takuya Nanami

Silicon neuronal network mimicking the olfactory nervous system of insects

Researcher

Takuya Nanami

Outline

The olfactory nervous system allows animals to utilize the odor information as a means to explore their foods, detect enemies, and communicate with their populations. In this study, we construct a silicon neuronal network mimicking the olfactory nervous system of insects. This study intends to provide a basis for utilizing the odor information and elucidating the information processing in the nervous system by the “analysis by synthesis” approach.

Koya Narumi

Printable, biomimetic softrobotic system

Researcher

Koya Narumi

Outline

We aim at implementing an automatic design tool to fabricate soft artificial muscles with predefined form factor and motion by placing elastomer which swells when stimulated by heat, using digital fabrication tools like 3D printer. We will also achieve biomimetic functions such as nerves for sensor and circuit and platelets to self-heal, by exploiting the property of the material used for the muscle.

Hideaki Hayashi

Probabilistic generative model of biosignals and neural networks for inference

Researcher

Hideaki Hayashi

Outline

In this study, we develop probabilistic generative models that describe the generation process of biosignals. We also propose neural networks based on probabilistic models for estimating such complex probabilistic generative models. The application of this study involves medical diagnosis support and human-machine interfaces using biosignals. Furthermore, this study leads to the development of a framework for integrating probabilistic generative models and neural networks.

Satoshi Funabashi

Achieving Tasks by Multi-Fingered Hand Using CNN and Distributed Tactile Sensor

Researcher

Satoshi Funabashi

Outline

When robots execute tasks like humans, multi-fingered hands with distributed tactiles sensors are necessary. Convolutional neural network (CNN) is used for controlling the hands by recognizing contact states with grasping objects. On the other hand, Each tactile sensor on the hands has different size and shape. Therefore, how to input this tactile information from the sensors is still an open issue. This proposal presents a method how to combine information from 3D distributed tactile sensors for training CNN. The uSkin sensors are mounted on an Allegro Hand and provide 736 tactile information and 16 joint angles in total. Since the sensors have 3D (x, y, z) vector tactile information, the input filters for x, y and z are prepared. The convolution layers are prepared for each tactile sensor mounted on not only phalanges but also fingertips of the hand. All the layers are concatenated resulting in building a tactile map based on the configuration of tactile sensors on the hand.

Junichiro Furukawa

Reproduction of human motion skill for exoskeleton robot control

Researcher

Junichiro Furukawa

Outline

The purpose of this study is to extract criteria from motion data including electromyography that can evaluate motion states of human, and clarify a method to adaptively control exoskeleton robots by utilizing the criteria. In particular, by expressing “know-how” of experts from motion data, this study proposes exoskeleton robot control that can evaluate the human motion and provide optimal assist strategy. This result will lead to a new motor learning support system by exoskeleton robots.

Moe Matsuki

Zero-shot Learning for Activity Recognition

Researcher

Moe Matsuki

Outline

Human activity recognition using sensor data like that of an accelerometer in a wearable device, needs a training dataset because this technology uses supervised machine learning. However, collecting a training dataset regarding all possible activities is a costly and tough job. To solve this problem, I propose a learning method which is able to recognize the activities which don’t occur in the training dataset. This method searches the relationship between the sensor data and text data which explain the activities. Then, I construct the estimated model to recognize the activities which don’t occur in the training dataset.

Yuto Miyatake

Probabilistic discretization methods for inferring parameters of differential equations

Researcher

Yuto Miyatake

Outline

When inferring parameters of differential equations, we first obtain a discrete model by discretizing the differential equations and then infer the parameters in the discrete settings. In many applications, however, the influence of the discretization is assumed to be sufficiently small without mathematical justification, and this could introduce a bias and lead to overly confident estimations. This study aims to develop probabilistic discretization methods to reduce such a bias, and to discuss reliability of the prediction obtained by simulating the inferred model.

Taiki Miyanishi

Improving Physical-World Question Answering Using Simulation and Domain Adaptation

Researcher

Taiki Miyanishi

Outline

Physical-world question answering (QA) using neural networks provides relevant text answers to questions about our daily life episodes. In spite of the fact that making physical-world QA datasets is crucially hard and has privacy issues, most existing neural QA models requires large amount training dataset to deliver superior performance. To overcome these problems, we use a life-simulator for making sufficient amount of training QA datasets. Moreover, we propose neural QA models having domain-invariant architecture with explicit memory components for bridging the gap between virtual & physical worlds and improving the physical-world QA performance. We evaluate our proposed cross-domain QA models using daily life episodes collected from a real-house and a life simulation game.

Tatsuya Yokota

Multiway Delay Embedding and Modeling of Tensors

Researcher

Tatsuya Yokota

Outline

In this study, we propose a new framework: A multiway delay embedding and modeling of tensors as a straight forward extension of the delay-embedding of time-series. The high-order Hankel tensor is constructed by multi-way delay embedding transform (MDT) and the Hankel tensor is represented by the low-rank tensor decomposition or manifold models. The proposed framework can be applied for the inverse problems by the following three steps: (1) MDT, (2) modeling, and (3) inverse MDT. The applications will be studied such as for image denoising, completion, or super-resolution.