[Information and future] Acceleration Phase Year Started: 2018

Yuki Arase

Elucidation of linguistic phenomena in paraphrases based on manual annotation and establishment of automatic phrase alignment methods for paraphrase collection

Researcher

Researcher image

Yuki Arase

Osaka University
Graduate School of Information Science and Technology
Associate Professor

Outline

A paraphrase restates the meaning of text using different expressions. It is crucial resource for natural language understanding and its applications, such as automatic question answering and multi-document summarization. The paraphrase is, however, not as simple as it sounds. It involves various linguistic phenomena; stating the same fact in different granularities, textual entailment, and so on. In this project, we analyze and annotate paraphrases to answer the fundamental question; what kind of linguistic phenomena make paraphrases possible. Furthermore, we develop methods to automatically identify phrasal paraphrases based on our annotations aiming to create a large-scale paraphrase dataset.

Satoshi Iizuka

Automatic Digital Remastering of Old Video and Image Contents via Deep Neural Networks

Researcher

Researcher image

Satoshi Iizuka

Tsukuba University
Faculty of Engineering, Informatin and Systems
Assistant Professor

Outline

I will establish a fully automatic architecture for remastering historical video and image contents via deep learning. Old photography usually consists of black and white images with noise and blur. Up until now, removing the deterioration and restoring the color required complicated image editing by an expert through trial and error. In this work, I will propose a novel architecture based on multiple convolutional neural networks to automatically detect time-related deterioration and restore them with color. I believe this will make people more familiar with historical image contents, which will improve the understanding of human history.

Nakamasa Inoue

Higher-Order Models for Semantic Concept Discovery on Multimedia Data

Researcher

Researcher-image

Nakamasa Inoue

Tokyo Institute of Technology
Department of Computer Science
Assistant Professor

Outline

The goal of this project is to create higher-order models for discovering semantic concepts from multimedia data. Here, a higher-order model is a model to generate detection models for unseen semantic concepts such as objects, actions, and scenes. We develop a learning framework to obtain a higher-order model from pre-trained deep neural networks. We evaluate the framework on a large-scale video dataset by finding and detecting unseen objects, event and scenes.

Akira Imakura

Development of a nonlinear nonnegative matrix factorization-based algorithm for deep neural networks

Researcher

Researcher image

Akira Imakura

University of Tsukuba
Faculty of Engineering, Information and Systems
Associate Professor

Outline

The backpropagation (BP) method, based on the stochastic gradient descent method, is the most successful and the de-facto standard algorithm for computing deep neural networks (DNNs) in several applications such as image recognition. In this research project, we develop a novel DNN computation algorithm and its parallel implementation using nonlinear semi-nonnegative matrix factorization (nonlinear semi-NMF) as a different approach from the BP method. We also develop matrix factorization-based algorithms for large data analyses.

Yuki Kubo

Construction of input systems for realizing Personal Interaction Space

Researcher

Researcher-image

Yuki Kubo

Nippon Telegraph and Telephone Coporation
Emproyee

Outline

In the near future, a user will see real world augmented with personalized information in anytime, anywhere through head-mounted displays such as a glass-type device. We call such information space Personal Interaction Space. To operate the information, flexible input methods are necessary to deal with information in all scenes. In this project, we develop input systems which utilize wearable sensors and devices to provide such flexible input methods.

Nahoko Kuroki

An interdisciplinary study of physical chemistry and information science for realizing CO2 free society

Researcher

Researcher-image

Nahoko Kuroki

Chuo University
Institute of Science and Engineering
Researcher

Outline

Efficient CO2 capture is indispensable for solving the global warming problem with keeping/improving our living quality. After selective CO2 absorption ability was found in some ionic liquids, chemists have tried to optimize the ability by changing the combination of ions that govern the physicochemical nature of ionic liquids. In this study, using an interdisciplinary approach of physical chemistry and information science, I will develop an efficient prediction system that can design ionic liquids with higher CO2 absorption ability. With a perspective in the field of information science, I would like to realize CO2 free society.

Ryota Kobayashi

Automatic analytic tools for event time series

Researcher

Researcher image

Ryota Kobayashi

Research Organization of Information and Systems
National Institute of Informatics
Assistant Professor

Outline

Despite the increasing demand for time series analysis, it has not been widely used due to the difficulty in developing a suitable mathematical model. A “good” model that yields useful results needs to be constructed carefully as the performance of time series analysis critically depends on the model. Herein, we focus on the event time series, i.e., the time stamp of event occurrence, and develop a machine-learning framework that can automatically analyze the event time series.

Maina Sogabe

Innovations for Data Acquisition in 4D Live Imaging

Researcher

Researcher image

Maina Sogabe

Kyoto University
Institute for Frontier Life and Medical Sciences
Program-Specific Researcher

Outline

Live imaging in vivo using the two-photon excitation microscopy is a new technology that is able to observe bio-activities deep inside living animals. In the future, the observation of large areas in the organs and tissues and the acquisition of higher resolution, high speed and long-term imaging data will be required to understand the precise behaviors in cells or molecules. The aim of this research is to optimize the live imaging systems by using information science and technology for the minimizing data size and imaging period.

Naoya Chiba

3D Measurements Based on Estimation Light Tranport Matrix between Camera and Projector

Researcher

Researcher image

Naoya Chiba

Tohoku University
Graduate School of Information Sciences
Master Course Student

Outline

3D measurement method on a projector-camera system is commonly used because of the popularity of 3D robot vision technology. However, there is still difficulty in measuring metallic objects, translucent objects, glass objects, or food items. Based on the Light Transport Matrix estimation method, we would like to perform the following: 1) try to achieve fast enough 3D measurement for industrial application, 2) to demonstrate robot vision application, 3) to recognize materials from reflectance specifications.

Kazutaka Nakashima

Connector aware shape decomposition for molding

Researcher

Researcher image

Kazutaka Nakashima

The University of Tokyo
Graduate School of Information Science and Technology Dept. of Computer Science
Ph.D. Course Student

Outline

Digital fabrication becomes popular in recent years, and Recently, a lot of innovative technologies for fabrication (e.g., 3D printing) emerged. Accordingly, in the field of computer graphics, many researches about fabrication are presented. However, most of them only optimize a shape and give it some functionality, and ignore cost for the actual fabrication. In my research, I also take such cost into consideration and optimize a whole of the fabrication process.

Yuko Hara-Azumi

Exploration and Design Framework of Architectures for Real-time Computation on Streaming Data

Researcher

Researcher image

Yuko Hara-Azumi

Tokyo Institute of Technology
Department of Information and Communications Engineering
Associate Professor

Outline

This proposal will realize novel multicore processors that achieve good trade-offs of performance-and-energy-efficiency and flexible customizability for a variety of embedded systems or so-called Internet-of-Things (IoT) edge devices, and build their design automation environment. The proposed processors enable real-time processing of streaming data that are increasing in the Big Data era, only within the IoT edge devices, leading to creation of new businesses/services and solutions for social issues.

Yusuke Matsui

Quantized Linear Algebra: Fast and Memory-efficient Matrix Computation by Data Quantization

Researcher

Researcher image

Yusuke Matsui

The University of Tokyo
Institute of Industrial Science
Assistant Professor

Outline

We propose a fast and memory-efficient matrix computation scheme, Quantized Linear Algebra (QLA). QLA first compresses input vectors or matrices into short memory efficient codes. The mathematical operations over the codes such as the matrix product are efficiently computed using looking up techniques. By the proposed QLA, computationally heavy operations including large scale machine learning can be achieved using small scale computational resources.

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