« iepriekšējāTurpināt »
RESEARCH INSTITUTE (ATR)
The Advanced Telecommunication Research Institute (ATR) is described.
by David K. Kahaner
(MOF), and Ministry of Posts and independent. Practically, though, ATR BACKGROUND
Telecommunications (MPT)], although can be considered as part of the NTT
ATR is viewed as the major laboratory family. I first visited the Advanced Tele- associated with MPT (the Electrotech- The four ATR laboratories are cocommunication Research Institute nical Laboratory (ETL) is associated located in a building in Kansai Science (ATR) more than 18 months ago, shortly with MITI even more directly). Actually City, a new area about halfway between after its 16K node Connection Machine JKTC provides only 70% of the funds Kyoto and Osaka. I was told that buildCM-2 had been installed (see my reporting for the four laboratories (or proj- ing costs were about $70M. Inside the “The Advanced Telecommunication ects) that represent the science com- ATR building there is a fifth laboraResearch Institute (ATR),” Scientific ponent of ATR. The laboratories were tory (called NTT Communication SciInformation Bulletin 15(4), 4-5 (1990)). established as 7- to 10-year projects; ence Laboratories (CSL)], which started My current visit was hosted by two will expire in 1993 and two in 1996, operation in July 1991. This operation
although there is every reason to think is not part of ATR but is just leasing the Dr. Kazunari Nakane
that they will be renewed, perhaps building space. Right behind ATR, NTT Head, Cognitive Processes Dept with slightly different emphasis. NTT owns land, where they may establish a ATR
and 140 other Japanese companies have new building for this laboratory. If this Auditory and Visual Perception Lab also provided private investment and happens, it will be the only official NTT Seika-cho, Kyoto 619-02, Japan the four laboratories get the remaining laboratory outside of the Kanto area. Tel: +81-7749-5-1411
30% of their funding from this. The Almost all the researchers now at CSL Fax: +81-7749-5-1408
same companies also fund ATR Inter- were transferred from NTT laboratories E-mail: firstname.lastname@example.org national, which supports the labora- in Musashino or Yokosuka in the Tokyo
tories and provides research facilities area. who also provided a great deal of assis- and support structure. ATR Interna- When I first visited in 1990, ATR tance in the editing of this report and tional has a president and board of was the only building in sight. Now it is whose patience is sincerely appreciated. directors. It is jointly owned by NTT surrounded by others in various stages
The ATR organization is compli- and the 140 other Japanese companies of construction, including several cated. It was established in October with NTT holding about 60% of its hundred residential townhouses. Across 1985 as a result of the privatization of stock. To make it more complicated, the street a large Sumitomo is going the Japanese telephone company. At the three key people at ATR Interna- up. Across from that, a big project that time the Japan Key Technology tional are retired from MPT, the regional called the Keihanna Convention Center Center (JKTC) was set up, funded mostly Kansai government in which ATR is is under construction. (KEI-HAN-NA by the dividends from NTT stock that is located, and NTT. Each of the individ. is a combination of the abbreviation of owned by the Japanese Government. ual laboratories has a president: two three words, Kyoto, Osaka, and Nara.) The Government now owns 60% of the are from NTT, one from KDD, and one It will be a convention center with stock of NTT. JKTC funds about 40 from MPT. In addition, many of the hotel and so on. Those readers who projects, of which by far the largest is staff members within each laboratory have been to Japan should imagine ATR.JKTC top management includes are part of a particular NTT organiza- what Tsukuba Science City (near Tokyo) representatives of various ministries tion. Thus there is a great deal of verti- must have looked like during its earli[Ministry of International Trade and cal information flow, and for similar est days to get a sense of the area around Industry (MITI), Ministry of Finance reasons the laboratories are rather ATR. Like Tsukuba in those days, many ATR staff members feel isolated from easily be absorbed back into a sponsor- coming to ATR is a kind of stepping more established parts of Japan, but ing company, a university, or a research stone for them. And, in general, many this will change as the community institute when the ATR projects end. researchers have been able to make develops roots.
Staff experience is also an issue. As excellent use of the experience, results, Almost half the ATR management is typical of Japanese companies, there
is typical of Japanese companies, there and ideas, etc., from ATR and, in some are on rotational assignment from a are relatively few employees with Ph.D. cases, were placed in very good posicouple of NTT's laboratories or divi- degrees. Companies that support ATR tions. This is especially true for NTT. sions. And the other half are mostly clearly use the laboratories as a place On balance, a few years at ATR seems from laboratories of KDD (interna- to put younger employees for training. to be a good thing for the scientists tional telephone company), NHK (semi. There was some difference of opinion involved. national radio and TV corporation), among the Japanese I spoke to as to The four laboratories are mostly and CRL (Communication Research whether an assignment at ATR was in independent, although they are housed Laboratory of MPT). Most of the staff anticipation of good work to follow in the same building. The independence members are also rotational, coming (back at the home company), a reward, is enhanced by giving each laboratory for a few years. About one-fourth are or a banishment. This is very company its own president as mentioned earlier. from NTT, but other Japanese com- dependent. However, most companies This was a bit strange to me, as the total panies are also represented. There are send their good engineers because they staffing of 260 is low enough that there also visiting Western researchers. I met are representatives of the company and could be a great deal of interaction, four, three Americans and a Frenchman. they want them to look good.
although each of the laboratories is Their appointments were temporary, Researchers who come with defi- aiming at its own specific research goals. 1 year at a time, but apparently could nite ideas about what to work on seem However, given the complicated funding be extended several times by mutual to fare quite well. I was told that each arrangements, it is not too surprising. agreement. Each of the four laboratories and every researcher knows what he A good example of the indepenhas a few senior scientists who are either will do at ATR because there are nego- dence is the use of the Connection permanentor at least on very long term tiations on this subject before he joins. Machine. This was purchased primarily assignment. These people provide the However, there is the usual trial and for the computational studies of cognitechnical leadership and continuity. error associated with research topics. tive processes carried out in the Audi
Rotation of staff in and out of the Perhaps Western researchers have even tory and Visual Perception Research laboratories provides a collection of more flexibility in this matter. I Laboratory. While it is extensively used, new ideas but also leads to some under- believe, though, that many of the younger I was told that some time is available utilization of equipment that is ordered Japanese are assigned to projects. Most (there might be some differences of for use in one near term research proj- new staff will work on a project sug- opinion about this), but researchers ect but is of less interest to the next. gested by an advisor or department from other laboratories and even other However, even in such cases there is head, by joining an existing project, or
head, by joining an existing project, or groups within this laboratory are occaresidual benefit, as the overall research by doing research alone. There is also sionally discouraged from using it. Dr. projects are generally well planned and the problem of what some of these Nakane explained to me that funding are all heading toward some very long Japanese will do when their ATR assign- for the machine is for specific research term goals, with natural near term ment is complete. Of course, most will jobs (rather than for other purposes or revisions and adjustments. I was told return to their companies, but several mere interest). This requires careful that most equipment will be used in have gone on to universities. One accounting and designation of users. succeeding projects. However, staff expressed concern to me that the basic Still, such an important resource ought turnover is definitely ATR's major research he was doing at ATR might to be of great help to scientists throughorganizational concern, and I was told not be supported in his home labora- out ATR. that there are new plans to develop tory and hoped that he would be able to Laboratory staffing is as follows. In more permanent positions. This is a find a good spot when he returned. On 1990, total staff was 262 and total buddifficult issue, because the ATR labo- the other hand, some returning get was Y7.5B, about $50M. This does ratories are not permanent. If ATR employees from Sony, Matsushita, and not include the space occupancy fee hires any permanent employee they have others have been known to get over- that ATR International charges to each to provide for the employee's long- seas assignments (e.g. to the United of the research and development (R&D) term career plan. I was told that one States, Germany, the United Kingdom, organizations. I mentioned this figure approach is to hire only those that can etc.) after their ATR assignments. So to a colleague who had spent 3 years at ATR, and he felt that the budget figure Auditory and Visual Perception Labo- • Advanced antennas, active array was severely underestimated, perhaps ratory (AVPL) (E-mail: yodogawa@atr- technology for mobile antennas, because of the complicated funding hr.atr.co.jp). ATR publishes the ATR methods of mitigating multipath situation. He thought that a more real- Journal, which is mostly in Japanese. propagation problems, microwave istic figure was twice that.
However, they also publish annually a circuit integration, signal processing.
bound collection of the staff's techniName of Laboratory:
cal reports. Many of these are in English, • Optical and electronic devices,
and all have English titles and abstracts. growth and characterization of Communications Systems
37 The four laboratories and their pri- semiconductors with precisely conInterpreting Telephony 52 mary activities are as follows.
trolled atomic configurations, nonAuditory & Visual Perception 57
linear optical devices. Optical & Radio Communications 40 Communications Systems:
Main goal: Human-oriented intelligent
Main goal: Automatic interpreting Invited international researchers 24
telephone Invited domestic researchers
15 • Communications with realisticsenOther researchers
sations, automatic three-dimensional • Speech recognition and synthesis, Permanent staff researchers
9 (3D) shape acquisition, recognition, speech database.
comprehension, modeling, manipOther Staff:
ulation, and display.
• Interface between speech and lan
guage, spoken language processing, Assistants
42 • Nonverbal interfaces, recognition knowledge base, speech and lanClerks
28 of facial and eye-gaze directions, guage integration. ATR International
6 understanding gestures and hand
movements, integration of visual and • Machine translation, grammar for Publications and patents have been speech information.
analysis of Japanese, dialog intergrowing since ATR's founding.
pretation, contextual processing. • 3D image databases. Papers:
• Advanced dialogue processing, • Cooperative work environment for contextual processing. Year International Domestic design of solid objects.
AUDITORY AND VISUAL 1986
45 • Automatic generation of commu- PERCEPTION LABORATORY 1987 21 282
nication software, extraction of real 1988 80 414
intentions, use of visual language to My visit was confined to the Audi1989 140 422
give specifications accurately, human tory and Visual Perception Labora1990 169 448
deliberative mechanisms in software tory, which has three departments. Its
design, knowledge base, easy to use main goal is an improved humanPatents:
specification description language. machine interface.
• Security, cryptographic techniques Visual Perception Department:
for large capacity (image) commu-
motion perception, binocular
stereopsis, character and word perOptical and Radio Communications: ception, image concept formation
• Optical intersatellite communica
tions, optical beam control, optical
• Brain activity measurement (eye
The figures above were given to me by
• Pattern recognition: neural network collaborating with almost a dozen • Measurement of EMG, articulatory
models for handwritten characters, younger researchers on an astonishing movement and acoustics for speech spatiotemporal patterns
variety of projects. This research has articulator and neural net model.
generated nearly 50 papers since 1987, • 3D object recognition
most in Western journals or in Pro- Computational model of visual cortices
ceedings of international conferences. and sensor fusion: Cognitive Processes Department: Kawato gave me a long list of the proj
ects that he is currently working on • Use of CM-2 to implement Markov • Cognitive processes for visual infor- (outlined below), but I had an oppor- random field (MRF) model of images
mation: neural computing models tunity to discuss only a few of these and learning of potential energies of pattern vision and spatial vision with his coworkers.
for hidden variables without teachers.
• Parallel processing mechanism: Learning trajectory control:
• Hierarchical MRF based on condimathematical analysis and synthesis
tional probability and application of neural net models, information • Learning control experiment of to pattern recognition. processing using neural net models rubbertuator Softarm by feedback
error learning neural net. Inverse • Shape from shading based on for• Learning and motor control mech- statics and inverse dynamics. Six- ward and inverse models of optics.
anisms: neural net models of human muscle arm model. Sensor fusion motor control, integrated learning for object manipulation.
• Integration of visual, somatosenof somato sensory and visual
sory, and motor control informainformation
• Feedback error learning for closed tion for object manipulation.
loop system. Models of regions of Hearing and Speech Perception cerebellum. Recognition of manip- There is a great deal of neural net Department:
ulated objects by motor learning. research in progress, including new
algorithms and applications to physi• Hearing mechanism: auditory Trajectory formation for arm and
Trajectory formation for arm and cal situations of speech and vision. The peripheral models, auditory prepro- articulator:
traveling salesman problem is another cessor for speech recognition
application being studied. The CM-2 is • Explanation of Fitts law by neural used heavily now for simulations. • Speech perception mechanism: co- net. Learning acquisition of forward (M. Hirayama showed me several very
articulation model and application dynamics of speech articulator on interesting demonstrations.) During my to speech recognition, speech recog- CM-2 using an electromyogram first visit, the ATR staff members were nition using neural nets, speech (EMG) as control input.
still learning about this machine and prosody
usage was low. I was told, however, that • Combination of cross validation the machine is now much busier. It is In this laboratory one of the most method and new information cri- extensively used as a neural network active researchers is
terion to estimate generalization simulator for the previously mentioned
capability of neural net models. studies of human perception, motor Dr. Mitsuo Kawato
Pattern generation and recognition control, and some other related topics. ATR (address as above)
using neural net.
Some of the computations/simulations Tel: +81-7749-51452
on CM-2 require an enormous amount Fax: +81-7749-51408
• Minimum-muscle-tension-change of machine use, say, 10 days of continE-mail: email@example.com and minimum-motor-command uous time, for example. Presently the
models based on biomechanical data CM-2 is working in "exclusive mode" Kawato's background is in biophysical of monkey musculoskeletal system but not in “TSS mode.” This is partly engineering and he was previously a with 17 muscles.
because of some problems with the faculty member at Osaka University.
TSS operating system (OS) and partly Kawato is one of the few semipermanent • Human arm movement using Opto- because of some decrease of compumembers of the research staff and he is track and recording EMG.
tational performance in TSS mode, which seem to decrease the efficiency the time integral of the square rate of substantial flow of visitors between ATR of the CM-2 usage. However, the new- change of the torque. Based on this and MIT. (Immediately after my visit, est version of TSS OS seems to have model he has developed a neural net- T. Poggio from MIT arrived to give resolved these problems and is to be work that reproduced Fitts law (some- several days of lectures.) installed very soon. However, it has times called speed-accuracy trade-off been used heavily enough so that of arm movement) and also used it to CONCLUDING REMARKS requests are being heard that it be study feedback-error-learning for illreplaced with a more powerful one. posed problems. Human sensory-motor My visits to ATR have been cordial The current machine was purchased control, such as arm movement, involves and revealing. About 10 scientists I on a 3-year installment plan and included a number of ill-posed problems. met were enthusiastically willing to 32-bit floating point units. ATR is hoping I think that it is unusual to find so describe their research, although the to get one with 64-bit floating point much biophysical expertise being put independent laboratory structure made units in the future, but at the current to use in modeling, especially coupled it difficult for me to get a full picture. time this has not happened. (I believe with advanced computer hardware and My visits have been highly organized that there are three Connection software. In the area ofvision, there are and tightly time constrained. The laboMachines in Japan--one at ATR, a projects to automatically generate 3D ratory I visited is doing very excellent second at the Institute for Computa- images from a series of stereo-pairs basic research, but I still haven't seen tional Fluid Dynamics in Tokyo, and a and another fascinating project to the work on interpreting telephony or third at an undisclosed location, prob- analyze, quantify, and predict our communications systems in other laboably a private company.)
response to optical illusions. Projects ratories, and I would like to go back One of the most interesting aspects like these require careful eye-movement and learn about these. What I have of ATR's research is how tightly coupled measurements and distance monitor- seen is a group of capable Japanese, it is to biological models of real sophis- ing, as well as a deep understanding of supplemented by a substantial number tication. A significant aspect of the current ideas in cognition. These are of long-term Western visitors. There work seems to be associated with difficult mathematical problems, too. are still some birth pains both at ATR modeling and then verifying models Our brain essentially solves the inverse and in the surrounding area; ATR is with complicated physical measure- problem of reconstructing a 3D visible barely 5 years old. Finally, procedures ments. One example of this is related to surface from the data of a two- need to be established to insure that understanding the mechanism behind dimensional (2D) image projected onto ATR is not only a training ground for the articulatory to acoustic transform, the retina. (Solutions of this problem young Japanese scientists and a tempoi.e., to understand the process by which are related to extensions of regulariza- rary or sabbatical station for Western linguistic intentions become speech. tion techniques, some of which are very
tion techniques, some of
which are very ones but also to insure that ATR will Speech production entails extraordi- well known in computational mathe- develop the continuity and world-class nary coordination among diverse neuro- matics.) There certainly is plenty of stature that will enable it to accomphysiological and anatomical structures. hardware available for experimenting. plish the major long-term goals it has These unfold through time to produce While I was there two technicians from set for itself. a complex acoustic signal that conveys Canada were visiting ATR to help set to listeners something of the speaker's up a pair of Optotracks, sophisticated intentions. ATR's research has used distance measuring equipment costing measured movements of the articulator in the neighborhood of $60K each. (upper and lower lips and horizontal Workstations are everywhere, along with and vertical jaws) and EMG data from associated peripherals. The ATR buildfour separate muscle groups and used ing and laboratories are large, spacious, these to train a neural net to generate and very well equipped. One of the motor commands to the articulators. only places in the United States that I
Another especially elegant exam- know of, where comparably subtle work ple is work by Kawato on studying the is going on, is at the Massachusetts trajectory followed by human arms. Institute of Technology's (MIT) brain Kawato proposed that this minimizes research department, and there is a