Artificial Intelligence Research Center

03
 AIRC center director

Junichi Tsujii

S martphones, cleaning robots, self-driving cars…the move

towards to incorporating artificial intelligence (AI) in all aspects
of society is accelerating. While there have been a number of AI
problems more effectively.

J apan has a number of strengths with regard to AI research.

One is its technological foundation in manufacturing. For
booms in the past, they were limited to laboratory research. Finally, example, we would have many advantages in research on self-
the time has come when AI is being integrated within society. driving cars since Japan has a few globally competitive automobile
AI has huge potential to help to solve various problems faced manufacturers which can share expertise with AI researchers.
by the Japanese society, which is the first country globally to Precision processing technology in the manufacturing industry is
experience advanced aging. It is important to note that the role of similar. Moving forward, it is highly likely that AI will revolutionize
AI is not simply to improve operational efficiency or to lower costs. the manufacturing workplaces which are the pride of Japan.
Rather, AI will also take on the role of improving the quality of Japan also offers top-class global services in the fields of
services in various industries. healthcare and caregiving. If the huge amounts of previously
For example, in the healthcare field, AI is capable of analyzing accumulated data and expertise can be shared using the latest
lung x-ray images and detecting signs of lung cancer that would technology, including big data analytics, it will become possible
escape human observation. As another example, AI is suited to to offer advanced services in even more locations. This will
quickly detecting problems that could affect the stable operation of result in possibilities to move towards healthcare plans that are
power-generating windmills. In other words, AI is highly effective increasingly personalized for the individual.
in detecting phenomena that have a tendency to be overlooked Furthermore, it must not be forgotten that Japan has cutting-
by humans. On the other hand, there are still many phenomena edge research environments for science and engineering. If
in society that can only be understood by humans. The “quality” the enormous volumes of scientific knowledge collectively
of intelligence possessed by humans and AI is fundamentally possessed by researchers all over Japan including businesses and
different. The future of AI can be considered the exploration of universities, can be gathered in the form of data and linked with AI
methods to combine the two types of intelligence in order to solve systems, it will surely bring about new societal innovation.

02 NATIONAL INSTITUTE OF ADVANCED INDUSTRIAL SCIENCE AND TECHNOLOGY ( AIST )

 We want to create a “space” for
open innovation that connects together industry,
academia, and government.

F undamental research into AI is advancing rapidly via

rigorous efforts worldwide. This research extends to many
areas, including machine learning, simulation technology,
these efforts, with our vision being to encourage collaboration
among companies. We believe that creating a “space” where
manufacturers from various industries can collaborate even if only
natural language processing, and development of computational on fundamental technology, would be of huge benefit in improving
architecture for AI. Japan also has excellent research organizations AI technology in Japan.
in industry, academia, and government. We believe it is our
mission at AIRC to offer a “venue” for open innovation that
connects the proprietary data and expertise in each of these areas.
In the US, a leader in AI research, data and expertise are collected
by private companies, and there is a fierce competition to acquire
superior engineers. Because this is the current reality, we would
like to present a model to the world of how a public organization
coordinates AI technology. Only a public organization can take
on the role of promoting the sharing of data that cannot be made
public by businesses and universities. In addition, cooperation with
overseas research organizations is essential.
To this end, we want to actively increase the number of foreign
researchers that we employ and expand our horizons to carry He graduated from Kyoto University Graduate School in 1973,
our joint research with foreign organizations. Naturally, we are PROFILE obtaining a PhD in engineering. He became Associate Professor
at Kyoto University in 1973, Professor at Manchester University
also putting effort into collaboration with domestic businesses. Junichi Tsujii in 1988, Professor at Tokyo University Graduate School in 1995,
Currently we have formed a number of joint research teams and Principal Researcher for Microsoft Research Asia (Beijing) in
2011 before taking on his current position in 2015. He still holds
with manufacturers, but we would like to further accelerate a Professorship at Manchester University.

03
 Co ncept
・AI that can be embedded in the real world

・AI that solves problems through cooperation with humans

・AI that can explain things

The goal of research at AIRC is “AI that can be embedded in the real world.” In other words, the goal is to achieve AI that
solves complex problems in society, in cooperation with people in the real world. We ultimately aim to implement AI in
society in a diverse range of sectors, including the service industry and healthcare/caregiving services, in order to help
boost industrial competitiveness and achieve a more abundant society. Accordingly, the role of AIRC is to serve as a core
hub for promoting large-scale research, in collaboration with eminent and up-and-coming researchers in AI and related
disciplines from Japan and worldwide.

R esearch strategy
Field

Ⅰ
Mobility

Field Field

Ⅱ We are developing next-generation technology

for achieving freedom and safety in movement. Ⅲ
Productivity Health, Medical
Manufacturing industry
Care, and Welfare
Service industry

AIRC
We aim to create new services by supporting We are developing diagnosis and drug discovery technologies
manufacturing workplaces. so that people can enjoy long, healthy lives.

Field Field

Ⅳ Ⅴ
Security and Other
Safety Fundamental technology
Infrastructure

We are using AI to enable automated explanation in We are building an open innovation platform for
videos and for evacuation guidance during emergencies. industry, academia, and government.

04 NATIONAL INSTITUTE OF ADVANCED INDUSTRIAL SCIENCE AND TECHNOLOGY ( AIST )

 Field

Ⅰ Mobility

Utilizing AI in “global monitoring of the Earth” and “smart mobility”

After the time of motorization in the 20th century, human beings from the tremendous satellite imagery by automatic detection of any
are searching for new forms of mobility with more “freedom and kinds of objects and events on the Earth.
safety”. To this end, we build a multiscale framework to integrate all [2] Smart mobility — While self-driving car technology advances,
the geospatial information collected by diverse mobile agents, such coexistence with human drivers requires more “human-like”
as Earth-observing satellites, aircraft, drones, and the self-driving understanding of the surrounding situations. Self-driving agents
cars/robots. The raw geospatial data will be converted into semantic must grasp accurate geospatial context in real-time through
maps to support efficient navigation of the mobile agents. In this semantic mapping of their total environment. Research is being
field, there are the two main research areas, i.e., “global monitoring conducted to support their autonomous guidance and control by
of the Earth” and “smart mobility.” constructing dynamic 3D map.
[1] Global monitoring of the Earth — The innovative downsizing of In this field, technology is being developed in which machine
Earth-observing satellites has realized “constellations” consisting learning is used to detect terrestrial features accurately from
of hundreds of micro-satellites. The resulting huge amount of Earth satellite images, as well as technology for automatically creating 3D
imagery must be automatically analyzed by AI instead of human maps using roving automobiles and autonomous mobile robots.
eyes. In order to promote the commercialization, we are developing
an intelligent system which efficiently collect beneficial information

Theme Theme
Automatic detection of terrestrial features Automated 3D map generation from roving
1 using satellite images
2 automobiles
Using machine learning and deep learning technologies, huge amounts of New automated 3D map generation method is being developed that makes
imagery received from Earth-observing satellites are analyzed in a fully geometric registration of new observation points with the reference　2D map.
automatic way. Not only will this enable specific features, such as solar This technology reduces accumulated “distortions” in measurement data,
power plants, to be identified, it can also help in detecting and understanding thereby producing more accurate 3D map. This is expected to be utilized for
general changes from long-term satellite observations. the commercialization of self-moving agents.

Initial 3D map produced by roving Georeferenced 2D map

robots (Geographical Survey Institute)

©2016 Digital Earth Technology, Digital Globe

Corrected 3D map after the registration
with 2D map

Note: This research is supported by the NEDO (New Energy and Industrial Technology Development Organization) Note: This research is supported by the NEDO project, “Core Technology Development of Next-generation AI and Robots.”
project, “Core Technology Development of Next-generation AI and Robots.”

Development of the mixed-reality platform in which spatio-temporal structures of physical space is authentically
Future
transported to virtual cyber space
prospects
Development of self-driving cars equipped with prediction systems of other vehicles' movement

05
 Field

Ⅱ Productivity

Manufacturing industry

Improvement in quality and precision of manufacturing

The adoption of AI is moving ahead even in the workplaces for simulated factory will be constructed that adds not only the situation
manufacturing that is the pride of Japan. but also processing procedure, as we work on understanding the
In mass production workplaces to date, machines with recorded operational state of moving machinery and failure prediction.
operation procedures have been called on to repeat the same In this field, we have entered the verification testing stage using
movements. Moving forward, as lower volumes of production for a prototypes in the development of a system that detects signs of
greater variety of products become more common, the technology failures and danger in wind power-generating equipment, and that
nurtured in these environments will begin being applied in other enables predictive maintenance. We are also developing technology
areas such as agriculture and the home. As this happens, machines to enable robots to handle objects without form such as towels or
will be needed that can respond to the situation more flexibly, and clothing in which their shape cannot be modeled beforehand. Here,
that can optimize the system. the latest technology is used including imitation learning methods
To this end, technology will be needed that can automatically that utilize deep neural networks.
generate robot work movements. Research is being conducted on The potential of AI is moving beyond the role of achieving greater
robots that can implement automated work process planning by efficiency that has been the mandate of machines heretofore, and
presenting the final state of an assembly procedure. In addition, a onto improving quality and accuracy.

Theme Theme
Detecting signs of failures in wind power
1 generation
2 Imitation robot

A system is currently in development that uses machine learning to We are developing "imitation learning method" that enable to learn the
detect signs of equipment failure that would be difficult to detect from relationship between visual information and teaching motion, through end-
visual inspection. Currently, sensor data is being collected from 41 actual to-end neural networks. And the method will enable motions whose teaching
domestic large wind turbines, and large-scale verification testing is being was difficult, such as picking up unknown objects, and folding of soft
conducted. The goal is for a paradigm shift from responding after the fact to objects. A “teaching-less” robot is definitely not just a dream.
maintenance that prevents problems and maintains safety.
Achieving stable and
Machine learning is conducted on normal operating repeatable folding motion
Installation of various sensors (IoT) states under various conditions, and deviation from of soft object
Vibration data these states is detected as an abnormality
Normal

Normal Quality in which

Sound data normal data is
state model consistent
Power Data that deviates
generation data Abnormality from normal is
judged to be
Feature abnormal
space

3200
Note : This research is supported by the NEDO project, “Core Technology Development of Next-generation AI and Robots.”
2700

2200
Degree of deviation

1700
Normal
learning Early July
1200 period Detection system: Abnormality detected
Workplace: No sense of abnormality
Mid-September
700 Workplace: Discussion
on response
200
Threshold
-300
2015.2.7 2015.3.29 2015.5.18 2015.7.7 2015.8.26 2015.10.15 Robot reproduces motion by watching a person move a box up and down
Note: This research is supported by the NEDO project, “Wind power generation advanced commercialization R&D/
smart maintenance technology R&D.”

Expansion of breadth of failure and danger sign detection, and improvement in prediction accuracy
Future
Establishment of key robot technology and greater integration
prospects
Creation of a robot that can appropriately respond to changing and complex environments and tasks

06 NATIONAL INSTITUTE OF ADVANCED INDUSTRIAL SCIENCE AND TECHNOLOGY ( AIST )

 Service industry

AI technology for transforming a product-based economy to a service-based innovation

In aiming for a society which creates new high-quality services that big data in prediction and control systems. Using these technologies
exceed existing concepts, it is essential to gather and analyze the as a foundation, we are focusing on efforts to reduce labor and power
experience and knowledge of people in the services industry and to requirements in simulated environments such as convenience stores
develop a framework for improving and reconstructing operations. through the use of AI and robots. In this field, we are also engaged in
Compared to manufacturing workplaces such as factories, service unique research such as the creation of an AI-supported framework
industry workplaces for caregiving, nursing, and education, as well that improves quality in service workplaces that create value through
as places that promote health and create community through hobbies collaboration between people with advanced knowledge and skills.
such as dance and music, are characterized by their large degree We will focus on how “non-tangible thing-centric” AI technology can
of freedom. For quality control of work in these places, there needs contribute to improving quality in the services industry. In addition,
to be a change in knowledge from “tangible thing-centric” to “non- there is also a movement to collect sensor data of healthcare and
tangible service-centric.” Specifically, research is being conducted caregiving facilities in living laboratories inside and outside the National
that gathers non-tangible things such as knowledge and awareness, Institute of Advanced Industrial Science and Technology (AIST), and to
and connects them to knowledge such as operational procedures and build an open database. We look toward living laboratories serving as
expertise in order to support greater operational efficiency. Research places for the development and verification of IoT sensing technology in
is also being conducted on using previously unobtainable operational service workplaces and living spaces.

Theme Theme
Improving quality of on-site services and Development of “living function resilient
1 supporting creation of new services
2 service” using living labs
Services with terminology and operational procedures that are heavily Living labs are open innovation environments for developing and evaluating
situation-dependent and differ from workplace to workplace will be AI that supports our everyday life in the real world. Life phenomenon data
standardized through ontological methods. This will increase compatibility are collected from living labs inside and outside of AIST (children’s hospitals,
between IoT and various AI apps, and lead to structure of operational senior homes, typical residences, etc.). We will analyze life phenomenon
knowledge and systematic collection of non-tangible things such as human related to human physical and cognitive changes, and build an open database
behavior, meaning, feeling, awareness, and body motion. so that this leads to the creation of new services.

Children’s hospitals Senior homes

Body motion quality measurement Site-driven knowledge discovery

Subjective Awareness Subjective
view view Greater
Quality efficiency
Feeling Thoughts
improvement Risk
Behavior reduction
Activity location Improvement location
Life phenomenon model
Rehabilitation hospitals Typical residences

Probability model
Non-tangible thing Data
Sensor Recognition result Behavior log

Site behavior measurement Life phenomenon ontology Living labs for observing and recording life phenomena related to physical and cognitive changes
Note: This research is supported by the NEDO project, “Core Technology Development of Next-generation AI and Robots.” Note: This research is supported by the NEDO project, “Core Technology Development of Next-generation AI and Robots.”

Improvement of services through change in perspective from a product-based to a service-based one

Future
Empowerment of services that use AI and robots
prospects
Accumulation of operational knowledge in wide array of service settings, and support of improvement and reconfiguration

07
 Field

Ⅲ Health, Medical Care, and Welfare

Leading-edge AI technology will change cancer diagnosis and drug discovery

Society is requiring the technology not only for medical treatment efficient algorithms, combinatorial factors causing genetic complex
but also for preventive healthcare in order that people can live a diseases become uncovered. Through collaboration with medical
healthy and long life. However, it is becoming the serious social organizations, we have entered a stage of verifying the effectiveness
problem that the heavy burden is placed on the doctors and of proposed technologies.
paramedics with knowledge and advanced abilities. In addition, we are also focusing on supporting drug discovery that
To overcome this situation, we are actively working on the research combines robot experiments and AI. Through AI and robot driven bio
of the medical engineering that will support doctors and improve research, automated searches of diagnostic markers and lower cost
diagnostic accuracy by using technology such as image processing individualized healthcare will become possible. This effort has been
and machine learning. We also challenging to develop the integrated recognized globally, and in the Multimer Predictions category of the
support system for cancer diagnosis, e.g. the computer-aided 12th Community Wide Experiment on the Critical Assessment of
detection system for breast ultrasonography. Techniques for Protein Structure Prediction (CASP 2016), AIRC was
In addition, through the development of new statistical methods with honored by being awarded 1st place worldwide.

Theme Theme
Computer-aided diagnosis in medical AI application to causal factor detection in
1 imaging based on machine learning
2 genetic disorders
The advances in image recognition technology enable the machine learning We are developing new techniques to understand the causal factors
system to support diagnosis in medical imaging. This technology is complex genetic disorders that to date have not been understood. Through
expected to contribute the great reduction of the burden on doctors and the the development of new statistical techniques with efficient algorithms,
improvement of diagnostic accuracy when it is used for screening and double it will become possible to uncover factors caused from simultaneous
checking. Specifically, a support system is currently under development genetic mutations. We will partner with medical organizations and conduct
that analyzes breast ultrasonography data and histopathological images to verification testing.
automatically detects suspected lesion .
Conventional method Only disorders from single factors can be detected
Candidate region
-log10P-value

The histopathological
image diagnosis support
software based on machine
learning displays areas
strongly suspected lesions
(cancer) by the red marker.

Our method Disorders due to multiple factors can be detected

(Diagram shows an example of 2 factors, but 3 factors or
more can also be detected)
Candidate combinations
-log10P-value

The breast ultrasonography

examination support
software based on machine
learning automatically
detects and displays the
areas suspected as tumor.

Supporting technology for diagnosis and decision of course of treatment, based not only on genetic information but
Future
also on examination information and lifestyle information
prospects
Drug discovery support that combines robot experimentation and AI

08 NATIONAL INSTITUTE OF ADVANCED INDUSTRIAL SCIENCE AND TECHNOLOGY ( AIST )

 Field

Ⅳ Security and Safety

AI visual recognition technologies of surveillance to protect safety of our society

AI technology is playing important roles to keeping safety, provide support in various fields.
security, and prosperity of our society. Video recognition and In addition, we are also applying technologies of large-scale
video explanation technologies provide accurate and ceaseless pedestrian tracking and simulation to support safe guidance in
observation, which are required to keep safety in factories, crowded area and evacuation planning in emergencies. These
hospitals, streets, and other locations in our society. technologies enable tracking of pedestrians by integrating numerous
Our research of video-to-text translation with high precision visual sensors and simulating behaviors of the pedestrians even in the
recognition will be applied for smart vigilance cameras. Most of case of a large event attended by 100,000 people or public spaces
conventional methods cannot achieve accurate recognition, because such as train stations or airport in real time. This is referred to as
they processed video information as a set of still images, not as sensing and simulation technology. AI is becoming an indispensable
a continuous time-series of images. Applying deep learning, our technology for safety of social systems.
system is able to understand objects in videos and events occurring
in videos accurately. In this way, AI’s power of visual observation will

Theme Theme

1 Video-to-text translation 2 Evacuation guidance using AI

Our AI system with deep learning technology generates sentences that We were able to achieve visualization of the overall movements of a large
explain the content and meaning of videos. Deep learning is utilized for number of pedestrians using real time pedestrian tracking technology and
recognition of objects and actions in video. Thanks to recurrent neural multi-agent simulation technology. Such visualization offers engineering
network (RNN), the system produces natural sentences for explanation. techniques to design evacuation guidance, especially in cases of large
numbers of people, whose overall movements have been difficult to
■ Our system can recognize complicated features, such as gender of recognize and have been controlled in ad-hoc ways.
person, detail of actions and so on.
Real time pedestrian tracking
technology that uses stereo
vision

Conventional method Conventional method

A man is drinking. A dog is playing with a dog.
Our method Our method
A girl is doing makeup. A boy is playing with a dog.

Conventional method Conventional method

A man is riding a car. A man is riding a bicycle. Sensing and simulation
Our method Our method technology that uses AI

A woman is riding a boat. A man is riding a bike. Note: This research is supported by the NEDO project, “Core Technology Development of Next-generation AI and
Robots.” and the JST (Japan Science and Technology Agency) CREST project, “Comprehensive Architecture of Social
Note: This research is supported by the NEDO project, “Core Technology Development of Next-generation AI and Robots.” Simulation for Inclusive Analysis.”

Future Connecting video explanation technology to specific applications in engineering, healthcare, and crime prevention
prospects Promoting societal use of pedestrian simulation

09
 Field

Ⅴ Other

Fundamental technology

Proposing new technology that supports and understands living scenarios

“Situational understanding” under various environments such as comprehends and supports living scenarios. In addition, for the
manufacturing workplaces, service industry workplaces, and the future we are exploring methods to apply this data to robot behavior
home is needed as a fundamental technology in the pursuit of co- planning.
existence of humans and machines. So AIRC is putting effort into In this field, we have established technology using deep learning
research of 3D object recognition, movement recognition, and that recognizes both shapes and functions of objects from 3D
environment recognition. data. In addition, we have built a cloud-based virtual reality (VR)
Specifically, we are constructing 3D data of objects found in daily environment and are accumulating huge amounts of interaction
life and movement data of people, and proposing technology that data between humans and robots.

Theme Theme
Object recognition technology using
1 Interactive robots based on cloud-based VR 2 deep learning
By building a cloud-based VR environment, a large-scale, long-duration We have successfully developed new object recognition technology using
interaction experiments using crowd sourcing have become possible. We deep learning technology that surpasses the accuracy of conventional
are building a framework in which hundreds of people simultaneously methods while using only a small number of observed images. We won first
participate in a 10,000-hour interaction experience, from which data can be place worldwide in the international 3D object retrieval competition SHREC
accumulated and shared on the cloud. 2017 for an algorithm developed for this research topic.

Conventional method ■ Deep Learning Robot

Method using actual robot
・Accompanied by hardware problems
・Large cost in gathering test subjects
・Difficulty in sharing interaction　
experiences
・Conversational experiment limited to
several hours

An image dataset of 12
categories of 132 everyday
items consisting of 21,120
Log in images was established
for training

Our method
Method using VR Our method
Conventional method
・No hardware problems
・Test subjects can participate without
regard to time or location
・10,000 hours of interaction experiences
90% classification rate 92% classification rate

can be shared on the cloud

80 observed images 20 or less observed images

Note: This research is supported by the NEDO project, “Core Technology Development of Next-generation AI and Robots.” Note: This research is supported by the NEDO project, “Core Technology Development of Next-generation AI and Robots.”

Infrastructure, etc.

■ AI Bridging Cloud Infrastructure (ABCI)

Building Global Artificial Intelligence COE
・Peak performance of 130 peta AI-flops or greater (floating point
We are building a high-performance, green, and cloud-based computing arithmetic performance at an accuracy suited for deep learning,
environment that can widely be used in industry-academia-government such as half precision floating point)
・Over 315 TiB of memory, over 2.7 PB/s of memory bandwidth,
collaboration for research and development of AI and IoT technologies. It will
and over 22 PB of storage
be built along with Global Artificial Intelligence COE that includes simulated
・Power consumption of less than 3,000 kW (projected annual
environments of the real world for data collection and proof-of-concept. In this average PUE of 1.1 or less)
way, we will accelerate the social-implementations of AI technology. SPEC

10 NATIONAL INSTITUTE OF ADVANCED INDUSTRIAL SCIENCE AND TECHNOLOGY ( AIST )

SPECIAL 6 TALK

Researchers Opening up AI Frontier

AIRC attracts the top AI researchers in the world.
Here is an introduction of what is currently happening
at one of Japan’s leading AI research centers.

Lead researcher

Yoichi Motomura
Motomura has served as director of the Japanese Society
SPECIAL TALK for Artificial Intelligence, the Society for Serviceology, and
the Behaviormetric Society of Japan. He specializes in next-
generation AI research and probability modeling. He is
Aspiring towards AI capable of specially appointed professor at Tokyo Institute of Technology,
01 mutual understanding with humans
guest professor at the Institute of Statistical Mathematics, and
guest professor at Tokyo University of Science.

■ Construction of industry platform using AI

Collection of data and Construction of phenomenon

knowledge from the real world models that people can understand
Data + Knowledge Phenomenon models

Learning and growth while

being implemented in society
Creation of computation models
Public systems that apply AI that can be easily processed by AI
Computational models
AI systems Computational models

Industry and living are being connected by AI technology, - AI will bring on the 4th industrial revolution
which is leading to big advances in business. What is this new
Implementing this technology in society is also an important mission of
age of AI in which manufacturing, distribution, and services
AIRC. The stages of AI technology implementation in society can be broadly
are being merged together? divided into the following three categories.
1:Greater efficiency in existing operations and activities = Use of AI in
- Next-generation AI understands phenomena behind big data existing services and methods.
2:Creation of new services = New services and methods made possible
It is said that the 3rd AI boom has now arrived. Driving this surge is a using AI.
“machine learning” approach, and big data created by the Internet and 3:Structural change in services and industry = Various services and
sensors have made this possible. Via machine learning, AI is steadily methods that use AI are dynamically connected together.
recognizing images, space, and texts in this world. Deep learning has We believe the path to this third stage as being the essence of the 4th
furthered advancement in image recognition, and space recognition from industrial revolution.
environment modeling is supporting self-driving technology. Natural While being implemented in society, next-generation A I will learn
language processing technology found in IBM Watson and Siri is expanding information from the real world and grow. The goal is to build phenomenon
the possibilities of text recognition. models that people can understand from workplace data and knowledge,
Furthermore, as we enter the IoT era in which sensors are installed in and these will be accumulated as computational models that can be
numerous devices and locations and connected via the Internet, there are easily processed by AI. By creating a platform for industry, deployment
expectations for AI to understand the phenomenon that lies within big data. and greater sophistication of AI, along with mutual collaboration will
An era has come in which AI has burst out from networks and phenomenon dramatically accelerate.
modeling has begun in the real world. If AI becomes able to recognize In this way if a dynamic connection is made between the Internet and real
not only conventional Internet usage data such as search words and world, or between the cyber and physical worlds, unprecedented business
purchasing history but also subconscious behavior and patterns in daily opportunities will be created. If this spreads to all industries, there is no
living and business, society is bound to change greatly. What is important question that future society will become even more prosperous. Various
here is modeling technology that enables AI to understand phenomena. areas of research are currently being conducted at AIRC including machine
This will require a model in which humans can understand phenomena and learning, probability modeling, service intelligence, and bioinformatics.
AI can perform calculations. This will be essential in creating AI that solves Here is an introduction of our researchers.
problems through collaboration with humans, which is the goal of AIRC.

11
 Research Report [ Seamless Map ]
Ryosuke Nakamura
Nakamura previously worked at the
Japan Aerospace Exploration Agency
SPECIAL TALK
(JAXA) before his current position.
His specialization includes planetary
Completely understanding time-space of science, satellite remote sensing,

02 this world
and satellite image processing using
machine learning.

the entire earth. As an effective method, automatic learning is now conducted

using deep learning techniques. Using pattern recognition technology, AI
will automatically infer the significance of image data from the input data of
satellite imagery and existing maps. This is a revolutionary breakthrough. In
addition, once inputting period-separated satellite image data of the same
area, AI will automatically recognize and explain what occurred over this time
period. For example, AI can provide information such as “The location that
used to be a school ten years ago is now a shopping mall.”

- From Odaiba to the world

Moving forward, the era would come where maps are updated daily, as
sensors equipped in self-driving cars worldwide record street-side data that
is changing every second. This map data will be viewed not on smartphones
or tablets, but rather on ski goggle-shaped wearable devices. While walking
Autonomous cars are becoming a reality in the 21st century. around town wearing these goggles, users might be provided information
Human beings are searching for new means of mobility. of nearby events by AI using satellite data, or be warned of observed
Technology that supports next-generation mobility using AI is abnormalities. This is probably an example of how the technology will be used.
being created at AIRC. However, it would be difficult for this type of innovative technology to be
quickly implemented in society. We would like to start testing in limited
areas such as a farm or seaport. We want to collect data focusing on the
- Recreating the real world inside computers Odaiba area where AIRC is located, and create a model for smart space
mobility technology. We believe that this area—which includes the driverless
The seamless map that we have been working on is not only a terrestrial Yurikamome train as well as the Tokyo Big Site exhibition hall and the National
map but also a cyberspace map that seamlessly covers everything Museum of Emerging Science and Innovation (Miraikan), where leading-
including indoors and subway station platforms. Our goal is to recreate edge science and technologies are on display—is ideal as a testing ground.
the real world inside computers, so we are trying to construct an efficient If possible, it would be great for AI to take on the role of guiding guests at the
mobility information platform from which the world’s entire time-space 2020 Tokyo Olympics and Paralympics. In addition, this technology is perfectly
construct can be understood. suited for applications in safety and security. Moving forward, we also want to
Underlying this platform is the use of terrestrial observational data derived put effort into research on AI that enables efficient evacuation guidance during
from artificial satellites, planes, and drones. Detailed city data derived from emergencies.
cars, autonomous mobile robots, and various indoor sensors is also added to To this end, we would like to actively collaborate with companies trying to
this database. It might be easier to understand by imagining zooming in on a create public infrastructure with new technology. By utilizing AIRC’s network,
Google Map’s satellite image and then switching to street view. Our goal is to we want to establish and spread a new mobility platform technology from
create this in 3D and connect everything ranging from inside the home to the Odaiba to the rest of the world.
department store basement. If the seamless map becomes a reality, it would
be a tremendous help to the mobility not only of people but also robots and
self-driving cars.

- More simply recording the physical world

Everything above and below ground is currently being recorded moment

by moment to prepare for the advent of an autonomous car society. This
world has not only a space axis but also a time axis, so even recording the
same location every second generates an enormous amount of data. If we
had a technology that could analyze this image data and then categorize
and textualize time-space being viewed by people, it would be very useful.
You could say that we are searching for methods to more simply record the
physical world using leading-edge AI technology.
One example is a technology that categorizes specific terrestrial structures
by using deep learning. In the past, people used to look at satellite imagery The image on the left is an example of virtual space and real space being superposed and displayed. This
and then specify map information on the computer individually, such as roads, is displayed in a goggle-type wearable device as in the image on the right. By displaying images of the
same location superposed at different moments on the time axis, changes between the past and present
schools, and factories. However, this method would not be practical to cover can be understood.

12 NATIONAL INSTITUTE OF ADVANCED INDUSTRIAL SCIENCE AND TECHNOLOGY ( AIST )

 Research Report [ Machine Learning ]

SPECIAL TALK

03 Aiming for machine learning embedded in the real world

inference. Going into a bit more detail, I am applying Markov chain Monte
Carlo methods to a probability model using Bayesian inference in search
of new data analysis methods. Recently, I have been developing organic
catalysts jointly with AIST’s Interdisciplinary Research Center for Catalytic
Chemistry and Research Center for Computational Design of Advanced
Functional Materials (CD-FMat). By combining test results—using this
catalyst derives this type of chemical reaction and subsequently produces
this amount of substance—with molecular simulation results, we are
working to predict new catalyst reactions using machine learning.
Asoh: What kind of research are you engaged in, Hayashi?
Hayashi: Since my student days I have been studying neural networks
and data analysis. Currently, I am doing research in data analysis methods
using tensor factorization. This is a method of machine learning that uses
the idea of matrices in mathematics, and can be regarded as a method of
compressing huge amounts of data. For example, a chart that shows how
The machine learning research team is comprised of Facebook users worldwide are interconnected can be expressed as a huge
researchers in various areas of specialization including matrix such as 1 billion x 1 billion. The goal of my research is to make the
neural networks, Bayesian inference, and data mining. AIRC’s matrix simpler by finding “coherence” that reflects regularity within the
deputy director spoke to two young people at the front lines of data.
Asoh: It is very important to find hidden grouping within data. Are you
machine learning research.
conducting joint research with companies?
Hayashi: I’m at the stage of exchanging information with AI related
- Conducting machine learning research in the place adjacent to venture companies, but I would like to use AIRC’s network to find a path
applications to application. Among research topics being pursued in the center, I think
“Understanding Cause of Genetic Disorders” (see page 8) is a good fit.
Asoh: AI and neural networks have previously experienced two booms in
the past, and now it is said that the third boom has come. We are engaged - What should researchers do in a world where machine learning has
in a broad scope of machine learning research including deep learning become ubiquitous?
at AIRC. I myself started research in neural networks around 1980 prior
to the second boom, and subsequently I have been conducting research Asoh: Do you have a future vision of what you would like to do at AIRC?
in machine learning using Bayesian inference and data analysis. Prior to Hayashi: Of course I would like to investigate tensor factorization
joining the center, what type of research were you two involved in? techniques, but as a researcher I am always thinking of what I can do in a
Nagata: I have been involved in research of theoretical approach since world where machine learning has become ubiquitous. Machine learning
my days as an undergraduate in the university. Specifically, I dealt with technology will undoubtedly spread into society as an element of intelligent
research that theoretically and mathematically analyzes how computer systems such as machine translation and self-driving cars. When that
algorithms work. In my doctoral studies in the graduate school, my happens, decision making that had until now been the domain of humans
research topic was searching machine learning methods that can be will be taken on by machines. In other words, situations might occur
universally used in natural sciences. I gradually wanted to do research in a where data generated from machine learning will be used by machine
location where the application of technology is emphasized, so I applied for learning. What would happen then? Although it may seem like a sci-fi
a researcher’s position at AIRC. movie, it is possible to imagine an attack from malicious AI. I believe it is
Asoh: There are researchers of various fields at AIST, and we have the responsibility of researchers like us to create robust and resilient social
an environment where we can work together with researchers from systems which can withstand such attacks.
companies, so I think it facilitates finding a path toward implementation Nagata: I also think machine learning and AI will be widely used in society.
of machine learning in society. By the way, what kind of research are you To this end, mutual understanding between AI and humans will become
doing now? more and more necessary. I want to strive toward creating a framework in
which humans can easily understand what machine learning is doing, as
- Machine learning has a wide variety including Bayesian inference, opposed to a feeling of AI being convenient but not well understood.
Markov chain Monte Carlo methods, tensor factorization, and so on. Asoh: This is precisely the realization of “AI that can mutually understand
with humans” and “AI embedded in the real world” that AIRC is striving for.
Nagata: Fundamental research on machine learning using Bayesian I’m expecting future success of you.

Hideki Asoh Kenji Nagata Kohei Hayashi

AIRC deputy director. He previously worked at He previously served as assistant He previously worked as a special
the Ministry of International Trade and Industry’s professor at the University of Tokyo researcher at the Japan Society for the
Agency of Industrial Science and Technology Graduate School of Frontier Sciences in Promotion of Science and as a specially
(currently National Institute of Advanced the Department of Complexity Sciences appointed assistant professor at the
Industrial Science and Technology (AIST)) prior and Engineering before his current National Institute of Informatics before his
to his current position. His specialties include position. His specialties include machine current position. His specialties include
neural networks and machine learning. learning and Bayesian inference. machine learning and data mining.
 Interview with Foreign Researchers

SPECIAL TALK

04 This is the home of mixed culture!

At AIRC, there are also many researchers from foreign databases to answer complex questions. The second is in “Recognizing
countries. What kind of future are they trying to create Textual Entailment” where our computers check whether what someone says
is correct. We have recently developed the best logic system for this task.
while engaging in various areas of research?
Nguyen: My major is in knowledge modeling and image/video analysis and
search. At AIRC, I am doing a research on detection of visual inconsistencies
- Attracted by leadership of AIRC director Tsujii within images and videos. This can be thought of as a support system for
making creative products.
Pascual: When I heard that Dr. Junichi Tsujii, who was my advisor during Seydou: I am doing more application-oriented research. Currently, with the
my graduate studies at University of Tokyo, had become the center director, NEC–AIST AI Cooperative Research Laboratory, we are working on deep
it made me want to also join the team. Dr. Tsujii is well versed in industry reinforcement learning and simulation to develop an AI control system that can
and academia, so I was confident he would provide a stimulating and unique stand up to unexpected situations. Specifically, I am aiming to build systems
research environment. that can work efficiently when faced with limited information or resources.
Kristiina: I also know Dr. Tsujii very well, since he was my academic advisor I am happy for the opportunity to work with highly skilled researchers from
during my doctoral studies at University of Manchester in England. When I industry and academia, and with varying fields of expertise.
learnt about AIRC and that it was recruiting, I immediately decided to apply. I Pascual: I also think that is one of the greatest strengths of AIRC. Many top
feel very proud to be able to work under his strong leadership.
Nguyen : I was about to complete my doctorate at Keio
University but still having many questions that I want answers.
After considering many options including remaining at university
to do research or starting my own business, I decided that the
best way to deepen my knowledge was to work with outstanding
people as a full-time researcher at a good institute. Coincidently,
I found AIRC’s recruitment, so I applied.
Seydou: After earning my doctorate at the University of Electro-
Communications in Tokyo, I wanted to continue doing research
in an industry oriented environment. This center, as a leading
Japanese AI research institution that connects industry and
academia, was the ideal place for me. Furthermore, research in
AI presents challenges and perspectives that I am excited about.
Nguyen: To be honest, I didn’t know much about AIST or AIRC
when I applied. Where did you hear about them, Seydou?
Seydou: I heard about AIRC from a friend who was working
at AIST’s Tsukuba Center, and then I researched more on its
website.
Nguyen: When I told my AIRC’s acceptance notice to one of my
esteemed advising professors at my graduate school, he was
pleased, saying, “You got the best job as post-doc!” I was very
relieved.

- Joint research possible with outstanding researchers from

industry and academia

Kristiina: Currently, my research concerns social robotics, and I

study the interaction between humans and robots. For example,
the development of the WikiTalk robot application allows users
to look for interesting information in Wikipedia via conversational
interaction with a speaking robot. The conversation model is
based on my CDM (Constructive Dialogue Model) approach which considers
P H O TO G A L L E R Y

interaction as cooperative activity between two agents.

Pascual: I am also interested in the communication between humans
and computers. Currently, my objective is to create AI that can understand
human language and reason about the world. I am mainly involved in two
projects. One project is on “Complex Q&A” in which AI uses large knowledge

14 NATIONAL INSTITUTE OF ADVANCED INDUSTRIAL SCIENCE AND TECHNOLOGY ( AIST )

 Pascual MARTINEZ-GOMEZ ESP Kristiina JOKINEN FIN Diep Thi Ngoc NGUYEN VNM Seydou BA SEN
His expertise includes She has experience as a researcher Her major is in knowledge His expertise includes
natural language at Nara Institute of Science and modeling and image/ deep reinforcement
processing, machine Technology and at Advanced video analysis and search. learning, optimization
learning and reasoning. Telecommunications Research Institute She is also interested in and algorithm design,
He also has research International. Her specialty is dialogue interdisciplinary knowledge and communication
experience in Germany. He modelling and multimodal human- integration. She is from networks. He is from
is from Spain. robot interaction. She is from Finland. Vietnam. Senegal.

researchers from industry and academia visit AIRC regularly and leading- trains, etc. However, the flexible time policy provided by AIRC has reduced
edge information is constantly shared. That is stimulating and motivating. In all the above stress. On the contrary, I often prefer not going home–because
addition, powerful computers, abundant resources and broad connections in doing research here is fun (laughs). How do you all feel about the work
industry and academia also support our research. atmosphere?
Kristiina: In my research of social robot interaction, cross-disciplinary joint Kristiina: The atmosphere is excellent. The research team I belong to has
collaboration is necessary. The high-quality research facilities and very active many interesting and thought-provoking discussions related to collaborative
and inspiring researcher network at AIST have a huge value for this. work. It is very exciting to be able to develop new technology while being
surrounded by researchers with a wealth of knowledge in different fields such
- Dynamic work-life balance and cheerful social gathering as ontology development, machine learning, knowledge engineering, pattern
recognition, and biomechanics.
Nguyen: I think AIRC is the home of mixed culture. People are smart and Seydou: The atmosphere is that you immediately feel as a member of the
friendly and with a very good sense of humor. Working here everyday is full team. It is a friendly work environment in which we share our activities and
of fun and self-improvement. I also appreciate the flexible work environment. interests during breaks, lunch time, and social gatherings. My research
I had always had a negative impression of work customs in Japan, which colleagues all speak English well, and they also bear with me with my
include arriving early for work, long overtime, commuting in fully crowded imperfect Japanese (laughs).
Pascual: It is easy to initiate contact with teams in other fields
within the center and conduct joint research. Researchers at this
center tend to be open and proactive. Since I came here, I made
many friends not only from Japan but many other countries, too.
We go out for lunch and dinner together and talk often. There
are a lot of restaurants in the surrounding Odaiba area.

- Future of AI research

Nguyen: My own view of AI is that it is for understanding

ourselves, in addition to AI technologies for people. Through AI
development, I want to address many important, yet difficult
questions such as what does it mean to see, or the meaning of
being a human.
Pascual: I also find interesting question of “what defines
humans.” M y long-term objec tive is to create A I that
understands humans and unleashes our potential. I want to
create a natural interface that supports decision making in
which AI operates as a brain using all of the knowledge of the
Internet.
Kristiina: My goal is to create interactive AI that can talk with
people, understand their behavior, assist them in everyday
tasks, and provide them with useful information for instance
about sports coaching, music instruction, and home caregiving
support. This kind of communicating robot agent can surely
make our daily lives more fulfilling.
Seydou: My goal also is to make people’s lives more fulfilling
with AI. By facilitating access to technology, it will surely be
possible to improve social welfare. In the future, I would like
to apply the expertise I gain from implementing AI toward
furthering development of my home country, Senegal.

15
 Interview with Young Researchers

SPECIAL TALK

05 Working in an environment that facilitates application research

Young researchers at AIRC also have their own missions and are - Presenting ideas during lunch meetings
actively engaged in research. We spoke to four of them about
Kanezaki: How does research demand from business and government arise?
their research in various fields.
Ogata: There are cases in which we propose joint research with businesses,
and other cases businesses come to consult with us. Recently in what is called
- Joint research with major companies concurrently in progress the third AI boom, offers have been rolling in from major corporations that
everyone is familiar with.
Yamazaki: My specialty is statistical theory, and recently I have been Nishimura: Joint research is flourishing not only with companies but also
engaged in its application to automobiles. Specifically, I am working on driver within AIRC. While there is division of roles with respect to keywords such as
modeling for a driver support system. Using in-vehicle sensors, we gather machine learning, probability modeling, and bioinformatics, there are many
data on velocity, information on accel and brake pedals, and steering wheel common points of research, so lateral connections between AIRC research
angle, and then analyze these data based on machine learning techniques. teams are also smooth.
The computer is made to understand driver psychological states and decision- Yamazaki: Innovation coordinators also assist us to bridge research teams. In
making processes under various road conditions. Currently, I am engaged in universities, research labs tend to be isolated, so I am thankful that we have
joint research with automobile related companies. When I decided to connect the lateral network here.
theory with application, I found AIST’s strong connections to industry was Nishimura: Often times, we enjoy mutual presentation of ideas during lunch
appealing. meetings in a frank atmosphere of “let’s have lunch together since we work in
Kanezaki: I also felt the importance of technology application, so I applied
for a research position at AIRC. Currently, I am primarily engaged in the
development of object recognition technology using deep learning (see page
10). 3D data acquired from Kinect and other sensors equipped in autonomous
mobile robots is analyzed to recognize object shape and categories. It is
something like 3D object pattern recognition using deep learning and I have
already started joint research with companies.

- Transforming service workplaces using knowledge engineering

Nishimura: My specialty is knowledge engineering. This is a research field

that garnered attention during the second AI boom. It can be thought of as
research that enables AI to use the knowledge of humans. Currently, I am
engaged in research to create a framework for improving operations in service
workplaces using AI. For example, in caregiving work, the detailed operations
vary in each workplace, therefore, it is difficult to create a common operation
manual. By incorporating a framework that systematizes knowledge centered
on workplace employees, I would like to help improve operations. I call this
method “knowledge explication.”
Ogata: Similar to Nishimura, I originally worked at the AIST Information
Technology Research Institute and joined AIRC when it was founded in 2015.
My work is to help transfer AI technology to workplaces in industry. At my
previous position I also did work to pave the way for applying technologies
such as pattern recognition, signal processing, and data analytics. My current
research centers on detecting signs of problems in wind power generation (see
page 6). This is the field of “smart maintenance” in which AI is used to detect
abnormalities, and demand for this from manufacturing workplaces continues
to increase.

16 NATIONAL INSTITUTE OF ADVANCED INDUSTRIAL SCIENCE AND TECHNOLOGY ( AIST )

 Keisuke Yamazaki Satoshi Nishimura Asako Kanezaki Jun Ogata
He previously worked as an He previously worked at She previously served as an He previously worked at
assistant professor at Tokyo the Human Informatics assistant professor at the the Information Technology
Institute of Technology. His Research Institute at AIST University of Tokyo’s Graduate Research Institute at AIST
specialties include Bayesian before his current position. School of Information Science and before his current position. He
statistics and traffic flow His specialties include Technology before her current has a doctorate in engineering.
analysis. ontology engineering, position. Her specialties include His specialties include spoken
knowledge engineering, and object recognition, machine language processing, and
linked data. learning, and intelligence robotics. acoustic and vibration analysis.

the same building.” For example, today a researcher from overseas presented Yamazaki: It is an important point to be able to share the latest technical
us her recent research. information. In fact, Nishimura and I have been engaged in heavy information
Kanezaki: Where does this take place? I’d like to participate in the lunch exchange.
meeting, too. Nishimura: That’s right. I will ask if Yamazaki’s probability model can be used
Yamazaki: Of course you are welcome to join. There are also many people with knowledge representation data I acquire from caregiving workplaces in
who join after seeing our lunch meetings and wondering what is going on. order to conduct analysis.
Lunch time is easier for people to get together. Yamazaki: The important point is “AI that can explain things.” I would like
Nishimura: There are many manufacturing engineers and researchers sent to work toward a situation where data acquired from machine learning can
here on assignment from companies. I think this is an attractive environment be interpreted so that AI won’t be a black box. In such a situation, there are
because there are people close by who have come in contact with various data absolutely no barriers of research fields.
gathered in their industry workplaces.
Kanezaki: I am primarily involved in fundamental research of deep learning, - Achieving greater sophistication while implementing AI in society
so it is a good opportunity to work with people who have more experience in
application research, and it is exciting to exchange ideas. Nishimura: Machine learning and deep learning are currently in the spotlight,
Ogata: At AIST’s Tsukuba Center, there are researchers not only of AI but but there are still many phenomena in the world that cannot be understood
also various other fields. Personally, I would like to see much greater research even after analyzing big data on the Internet. Operational improvement of
interaction across different fields. niche workplaces within the service industry on which I am researching is one
example. I think the ideal AI must be one that can be easily used by people in
the workplace and that can provide useful knowledge on the spot.
Ogata: I’ve been working with the team that focuses on technology application
and doing a more workplace-oriented approach. One thing I always keep
in mind is that AI may become meaningless unless it can be easily used by
ordinary people. As for failure detection of wind power generators on which I
have worked, we use machine learning to analyze data from actual equipment
at wind farms nationwide. I want to create a framework that reflects the views
of people in these workplaces.
Yamazaki: I think that is an ideal cycle of increasing AI sophistication while
implementing it in society.
Ogata: I welcome an era when the image of AI being difficult to understand
is shed and when AI can be actively nurtured by people in the workplace.
Nishimura, do you have an ideal situation for AI in service workplaces?
Nishimura: From my perspective, it would be ideal for AI to be incorporated
into workplace operations and for it to naturally become more sophisticated as
a result of interaction. I also expect data and advice presented by AI systems
to change the behaviors of employees and improve operational flow. Kanezaki,
do you also have a vision for the future?
Kanezaki: I have an interest in machine autonomy. Specifically, I would like to
create an intelligence in an autonomous mobile robot using technology such
as deep learning. The history of human technology can be thought of as the
history of automating processes that humans can do. I would like to surpass
this and create AI that can think on its own and act. This is very possible using
today’s technology. I would like to achieve things that have not been thought of
before and give them back to society.
Yamazaki: I see, very interesting! I have long been involved in statistical
theory research, but jumped into the application research field. I would
like to further pursue theory in a place close to where technology is being
implemented in society. Fundamentals and application tend to be regarded as
different things in university research labs, but I would like to do both ends
P H O TO G AL L E R Y

at the same time. I think AIRC is the only place where both fundamentals and
application can be pursued in a leading-edge research environment.
Ogata: As someone involved in technology application, it is my intention to
connect people’s dreams to real life. To this end, I would like to see more
unique young researchers join our research.

17
 NEC – AIST AI Cooperative Research Laboratory

SPECIAL TALK
Washio: The third is cooperation between AI and autonomous systems.
Creating a new field that When self-driving cars become a reality, if all AI were to make decisions
06 combines simulation and AI
in a self-centered manner, traffic congestion could worsen rather than be
eliminated. I believe that research on “compromise” between AI is needed.
The research has not started yet, but I think algorithms developed here
can be applied to various scenarios in society such as in coordinating
the production activity of a company. A common point underlying these
research themes is to create algorithms with machine learning, while using a
simulation method in the cases of events or locations with limited data.
Tsuruoka: For example, there is the Japanese saying, “When the wind
blows, barrel makers make money.” It starts with the phenomenon of the
wind blowing, and through a series of inference steps, it eventually leads
to the conclusion that barrel makers make money. Inference based only on
symbolic knowledge is something fragile. In theory it could happen, but in
reality, it will not. Here, theory is augmented by combining a model that can
be realistically recreated. The idea is to augment theory with simulation.
Using this method, we would like to create AI that supports operations,
for example, at a plant. If AI were able to understand the details of plant
operation using simulation technology, when a problem arises, AI would
be able to explain why it occurs. In addition, it should probably be able
At the NEC–AIST AI Cooperative Research Laboratory, which
to present a solution, such as by indicating that “Recovery is possible by
was founded in June 2016, there are researchers not only following this procedure.”
from NEC, but also universities. What research is being Washio: The important point is it can provide an explanation. AI currently
conducted at this laboratory that features collaboration has aspects that are like a black box. “AI that can provide explanations” is
the goal of AIRC as technology that can provide the answer to the question
between industry, academia, and government?
“Why did that result occur?”

- Using Simulation to cover the areas with limited data - Aiming to apply “generic” methods in all social scenarios

Tsuruoka: The NEC–AIST AI Cooperative Research Laboratory is drawing Tsuruoka: Joint research with NEC is the first example, and there are many
attention as a place that aggregates the experience of company and companies approaching AIRC in this field. At universities, there are a number
university researchers together with the research expertise of AIRC. Dr. of joint research projects being conducted with companies, but I think it is
Washio, you have been at the lab since its inception in 2016, right? difficult to find a joint research project of this scale.
Washio: When I was invited to participate by AIRC, I was interested in the Washio: Currently, we are engaged in research with over 10 team members
theme of how to deal with various problems by combining simulation with per project. It would be difficult for not only a university but probably even
AI technology such as machine learning. And I was surprised to learn that for a company to achieve this scale. I think AIRC can be a major player in
private companies had a need for the areas with limited data, not for big data. this situation. At our research lab, we have many visiting researchers not
Tsuruoka: If we had to deal with the theme of analyzing big data with deep only from the University of Tokyo and Osaka University but many other
learning, we don’t need to stick to work at the lab with collaboration between universities as well. This has resulted in broadening the scope of research
industry, academia, and government. themes.
Washio: There are three main projects that are currently under way at NEC– Tsuruoka: From the perspective of researchers, the strength of AIRC is its
AIST AI Cooperative Research Laboratory. One is the convergence of machine attractive environment for AI application development. For example, with this
learning with simulation. This research project aims to understand complex collaborative lab, if there are excellent ideas, NEC will apply them in production
societal phenomena and to search for solutions to problems, by using workplaces or public projects with abundant knowledge and experience.
simulation technology to cover low probability phenomena such as disasters Washio: Our goal in this environment is not to solve niche problems of
and unknown situations, in addition to big data analysis of phenomena that businesses, but to develop generic and powerful methods and to apply them
occur on a daily basis in our society. This is an attempt to utilize engineering in all aspects of society. I think the NEC Central Lab also regards this place
solutions derived from this project to build livable cities and consider flexible not only as a company lab but also as a research base that is seeking lateral
countermeasures during emergencies. tie-ups in industry.
Tsuruoka: The second is my research theme, the convergence of automated Tsuruoka: I would like to invite young researchers and students, and make
reasoning and simulation. Through programming and inference using a simulator, this lab a place to develop talent for all industries.
we conduct a problem cause analysis and countermeasure deliberation, and Washio: I think there are very few research institutions even worldwide
present a solution in a format that people can understand and execute. The goal that aim to solve problems, at this scale, in the areas without much data, by
is to apply this method to a technology that supports operation of complex and combining simulation and AI technology. I would like to pursue research topics
dynamic systems, such as large-scale plants or IT systems. that nobody has thought of yet, and to produce world-first research results.

Takashi Washio Yoshimasa Tsuruoka

He completed his doctoral studies in the Department of Nuclear He completed his doctoral studies in the Department of Electrical Engineering and
Engineering at the Tohoku University Graduate School of Engineering. Information Systems at the University of Tokyo Graduate School of Engineering.
He was previously assigned as a guest researcher at Massachusetts He previously worked as a researcher at the Japan Science and Technology
Institute of Technology and a senior researcher at Mitsubishi Research Agency and as an associate professor at the Japan Advanced Institute of Science
Institute before becoming a professor at the Osaka University Institute of and Technology before becoming an associate professor at the University of
Scientific and Industrial Research. As an AIST special fellow, he is also the Tokyo Graduate School of Information Science and Technology. As a resident
18 Collaborative Research Lab Director. researcher at AIST, he is also Collaborative Research Lab Deputy Director.
We are also accelerating collaboration with private enterprises
and overseas research organizations!
ievement
Ach
AIRC wants to serve as a bridge for AI technology through

joint research with businesses. We have established AI collaborative

research labs with certain companies, and we are actively O t h er

seeking tie-ups with joint ventures.
ac t ivit ies

1 Business collaborations

Partnerships with large-

● We are joint research partners with over 40 private businesses.
2 scale public projects and
research organizations
● In addition to establishing the NEC–AIST AI Collaborative Research
Lab and the Panasonic–AIST Advanced AI Collaborative Research
Lab, we are also working toward ● SIP ( Strategic Innovation Creation Program)
opening a join t r e s ea r ch cen ter We have par tnerships in infrastructure maintenance and
with businesses. autonomous travel (dynamic mapping).
https://www.ith.aist.go.jp/ ● Other research organizations
We have a
joint research
par tnership with
the National Cancer
Stronger partnerships Center Japan.
3 with overseas research
organizations

● We are strengthening ties with overseas

research organizations such as University
of Manchester, German Research Center
for A r tificial Intelligence ( DFK I ) ,
C arnegie Mellon Univer sit y, and AI technology proliferation and
Toyota Technological Institute at 4 human resources training activity
Chicago.

●AI technology consortium http://www.airc.aist.go.jp/consortium/

●AI seminars and workshops https://airc.doorkeeper.jp/
● International symposiums, joint symposiums with Nikkei
● Venture outreach committee
● AI startup workshop
● Par ticipation in Next-generation AI and Robot Core Technolog y
Development Private Exhibition
● NEDO–AIRC–The University of Tokyo AI Talent Development Course
● NEDO–AIST joint symposium,“Core Technology Development of Next-
generation AI and Robots.”
 ACCESS

By train

■ 3-minute walk from the Telecom Center Station on the New Transit Yurikamome Line
■ 15-minute walk from the Tokyo Teleport Station on the Rinkai Line

By bus

1. Take the Toei Umi01 bus from the 2. Take the I No.30 or I No.32 Keikyu 3. Take the Mori No.30 or Mori No.40
Rinkai Line Tokyo Teleport Station or bus from the west exit of Oimachi Keikyu bus from the east exit of
the Monzen Nakacho Station on the Station on the JR Keihin Tohoku Omori Station on the JR Keihin
Tozai or Oedo subway lines Line. Tohoku Line.
→Get off at the Telecom Center Station stop, and walk for about 5 minutes.

From Haneda Airport

Take the monorail from the airport

→Transfer to the Rinkai Line at Tennozu Isle Station　→ Get off at Tokyo Teleport Station and walk 15 minutes

National Institute of Advanced Industrial

Science and Technology (AIST )

AIRC Artificial Intelligence Research Center

2-4-7 Aomi, Koto-ku, Tokyo 135-0064
E-mail airc-info-ml@aist.go.jp
URL http://w w w.airc.aist.go.jp/

2 017. 07