eugene@wilbur.nas.nasa.gov (Eugene N. Miya) (06/14/90)
To: Distribution
From: David Kahaner ONRFE [kahaner@xroads.cc.u-tokyo.ac.jp]
Re: The Advanced Telecommunication Research Institute (ATR)
12 June 1990
ABSTRACT: A visit to The Advanced Telecommunication Research Institute,
(ATR) on 15 May 1990 is described.
ATR was organized in 1986 with a combination of
government and private money for the following express purposes.
Plan and promote basic telecommunication research.
Establish and maintain research facilities and equipment to promote
integrated joint research opportunities for industrial,
academic, and governmental organizations.
Promote international technical research and exchange of
researchers.
About one third of its support is from private companies such as NTT,
while the remainder is governmental. In April 1990, ATR was
capitalized at slightly more than 51 billion yen, about $350,000,000 US.
There are four major laboratories within the complex. Their staffing is
shown below.
NAME OF LAB RESEARCHERS OTHER TOTAL
normal+invited
Communications Systems 30 3 33
Interpreting Telephony 38 7 45
Auditory and Visual Perception 45 10 55
Optical and Radio Communications 36 2 38
---- ---- ---
149 22 171
ATR Inst. Int. (admin) 6 25 31
---- ---- ---- ---
TOTALS 155 22 25 202
MAIN RESEARCH FACILITIES
Super minicomputers 15
Parallel supercomputers 2
Sound shielded booths 7
3-D image display equipment 1
High frequency network
analyzing system 1
Variable reverberation chamber 1
Radio anecholic chamber 1
Clean room 1
Last year the staff presented more than 500 research results via papers and
talks.
The main functions of the four ATR laboratories are as follows.
Communications Systems:
Communications with realistic sensations, automatic 3-D shape
acquisition, modeling, manipulation and display
Nonverbal interfaces, recognition of facial and eye-gaze
directions, understanding gestures and hand movements,
integration of visual and speech information
Automatic generation of communication software, extraction of real
intentions, use of visual language to give specifications
accurately, human deliberative mechanisms in software design
Security, cryptographic techniques for large capacity (image)
communications, secure telecommunication networks
Optical and Radio Communications:
Optical intersatellite communications, optical beam control,
optical modulation/demodulation
Advanced antennas, active array technology for mobile antennas,
methods of mitigating multipath propagation problems,
microwave circuit integration, signal processing
Optical and electronic devices, growth and characterization of
semiconductors with precisely controlled atomic
configurations, nonlinear optical devices
Interpreting Telephony:
Speech recognition and synthesis
Interface between speech and language, spoken language
processing, knowledge base
Machine translation, grammar for analysis of Japanese, dialog
interpretation, contextual processing
Auditory and Visual Perception:
Basic mechanisms of visual perception, character and pattern
understanding, scene analysis and understanding
Cognitive processes for visual information, parallel computing
principles, learning and motor theories of perception
Hearing and speech perception and recognition, auditory models
ATR publishes the ATR Journal six times each year. Although most of the
articles are in Japanese a few are in English, and often there are enough
figures, illustrations, and tables to get a sense of the topics. There
are also reports, whose titles (some in English) are printed in the
journal.
My own interest in ATR was specifically toward their research in parallel
processing and more casually in neural networks and graphical interfaces.
My main contact there was Dr. Noboru Sonehara, in the Auditory and Visual
Perception lab [sonehara@atr-hr.art.co.jp], and his boss Dr. Kazunari
Nakane, who heads the Cognitive Processes Department, [nakane@atr-
hr.atr.co.jp]. Both of them were extremely gracious, cooperative and
open with me, although it was clear that there was more than a little
sensitivity because of my association with ONR. Ultimately it was
suggested that half a day ought to be enough time for my visit. During
my visit many of the researchers were away, and I was only able to have
discussions with Nakane, Sonehara, and Dr. Masa-aki Sato in the Visual
Perception lab, Dr. Peter Davis who is working in optical materials, and
Kelley Picket who works for Thinking Machines and has been at ATR for
about six months.
ATR has a Connection Machine, [Note: this is the only CM in Japan]
although it has been modified to only permit
32 bit floating point computation. I am told that new trade agreements will
allow 64 bit chips to be installed. Picket who is responsible for
teaching the ATR staff how to use the machine, felt that it was slow
going to get some of the Japanese scientists to pick up the ball and run
with it, although Sonehara was one of his most enthusiastic users. At
this time the machine is not being used as effectively as it could be and
there are some internal questions about what groups are permitted access.
Only a few programmers are actually using the Connection Machine, although a
larger number of people are directing work by the programmers.
Most of the work on the Connection Machine is related to implementing
neural network models. For example, Sonehara sent me a paper of his
concerning the problem of converting a digitized image containing k grey
levels to one containing only black and white (k=2), so-called binary
representation of a grey level image. Unfortunately, the paper did not give
any details of the computational implementation. An earlier paper, using
analog methods was given by Koch, [Koch C., Marroquin J., Yuille A.
"Analog `Neuronal' Networks in Early Vision" Proc Nat Acad Sci USA, 83
pp4263-4267, 1986]. Sonehara also has studied interpolation methods for
image processing using an iterative neural network model which he also
implemented on the Connection Machine.
Dr. Sato described some of his work which has general application in the
area of speech synthesis. (A preprint is available from the author.
"APOLONN brings us to the real world: Learning nonlinear dynamics and
fluctuations in nature"). A major problem in synthetic speech is its
unnaturalness, due mostly to the lack of high harmonics, short waveforms,
and time dependent spectra. Sato and his colleagues are experimenting
with a nonlinear network, APOLONN, in an effort to solve this problem. In
the analysis of their network they are led to the mathematical problem of
minimizing an energy functional whose values depend on the solution of a
system of ordinary differential equations with prescribed initial
conditions. The analysis is quite elegant. But, as I have now seen
several times the computer simulation of this was done from the ground up
with no collaboration with a numerically trained researcher. The
integration is via fixed step Euler's method and this could surely be
improved. I cannot tell if such improvements would make an appreciable
difference in their ability to train the network to synthesize speech.
Dr. Davis is primarily working on optical devices. He has been studying
systems that can be made to behave chaotically. In his context this means
that they can have many configurations by making small changes in their
input. He is hoping that such devices can have applications as memories.
Although the techniques are quite different this is in the same spirit as
Michael Barnsley at Georgia Tech, who wants to capture the details of an
image in a few parameters and hence use them as a data compression
technique. In Davis' model of the devices he is studying are several
ordinary differential equations of "delay" or "retarded" type, wherein
the time derivative of the unknown depends on the solution at some past
times. I explained to Davis that there was new work on these equations,
and it is possible that some of his models will be used as test cases for
very new numerical solvers.
ATR is extremely well equipped, even in comparison to some of the other
industrial labs I have visited. If anything, it seems to be top heavy
with equipment and short of staff, although it is a very new facility and
is still in the recruiting stage. One of their well publicized projects
is to develop a "Dick Tracy" watch. As Seiko has just announced a wrist
watch with pocket pager there is plenty of interest in this research.
ATR is also working on manipulation in 3 dimensions. An operator wears a
specially instrumented glove and glasses while watching a computer screen
on which a 3-D computer generated image is displayed. The 3-D effect is
via the glasses. By moving the gloved hand in free space the operator can
effect the movement of various objects on the screen. Unfortunately,
none of the researchers on these projects was available during my visit.
CONCLUSIONS.
ATR is in a brand new laboratory in a new area, designated as Kansai
Science City, for which great things are planned. At the moment it is
pretty much out by itself, about one half hour train ride between Kyoto
and Osaka. Several square miles of ground around the facility have been
cleared to make room for other research and high tech commercial
buildings. A few of these are now going up. There will also be a new
international airport nearby. The handful of researchers that I met,
Davis, Sato, and Sonehara, are articulate and their projects are
interesting. They certainly have all the hardware they need. In many
large laboratories there are a few strongly directed people whose work
is well known and who provide vision and guidance. I did not know any of
the ATR scientists before I visited and I cannot tell if any fall into
this category. Many of the scientific staff are on rotational
assignments from industrial research laboratories and when they leave
ATR their projects often are left hanging. Perhaps this assignment is
viewed as a form of training by their parent companies. At the moment
the parallel computation facilities are being used as a research tool
for other problems rather than to study parallel computation. This may
change as the staff becomes more familiar with the equipment. There
does not seem to be any numerical analysis expertise or research being
performed there; it may not be needed. Nevertheless, ATR deserves a
much more comprehensive look than the afternoon that I spent there.
For information about ATR, contact
The Advanced Telecommunication Research Institute
Sampeidani, Inuidani
Saika-cho, Soraku-gun
Kyoto 619-02
Tel: 81 7749-5-1111, Fax: 81 7749-5-1108.
---------------END OF REPORT-------------------------------------------