Lapas attēli
PDF
ePub

ADVANCED TELECOMMUNICATION RESEARCH INSTITUTE (ATR)

The Advanced Telecommunication Research Institute (ATR) is described.

by David K. Kahaner

INTRODUCTION AND BACKGROUND

I first visited the Advanced Telecommunication Research Institute (ATR) more than 18 months ago, shortly after its 16K node Connection Machine

CM-2 had been installed [see my report "The Advanced Telecommunication Research Institute (ATR),” Scientific Information Bulletin 15(4), 4-5 (1990)]. My current visit was hosted by

Dr. Kazunari Nakane
Head, Cognitive Processes Dept
ATR

Auditory and Visual Perception Lab
Seika-cho, Kyoto 619-02, Japan
Tel: +81-7749-5-1411
Fax: +81-7749-5-1408
E-mail: nakane@atr-hr.atr.co.jp

who also provided a great deal of assistance in the editing of this report and whose patience is sincerely appreciated.

The ATR organization is complicated. It was established in October 1985 as a result of the privatization of the Japanese telephone company. At that time the Japan Key Technology Center (JKTC) was set up, funded mostly by the dividends from NTT stock that is owned by the Japanese Government. The Government now owns 60% of the stock of NTT. JKTC funds about 40 projects, of which by far the largest is ATR. JKTC top management includes representatives of various ministries [Ministry of International Trade and Industry (MITI), Ministry of Finance

(MOF), and Ministry of Posts and Telecommunications (MPT)], although ATR is viewed as the major laboratory associated with MPT (the Electrotechnical Laboratory (ETL) is associated with MITI even more directly). Actually JKTC provides only 70% of the funding for the four laboratories (or projects) that represent the science component of ATR. The laboratories were established as 7- to 10-year projects; established as 7- to 10-year projects; two will expire in 1993 and two in 1996, although there is every reason to think that they will be renewed, perhaps with slightly different emphasis. NTT and 140 other Japanese companies have also provided private investment and the four laboratories get the remaining 30% of their funding from this. The same companies also fund ATR International, which supports the laboratories and provides research facilities and support structure. ATR International has a president and board of directors. It is jointly owned by NTT and the 140 other Japanese companies with NTT holding about 60% of its stock. To make it more complicated, the three key people at ATR International are retired from MPT, the regional Kansai government in which ATR is located, and NTT. Each of the individual laboratories has a president: two are from NTT, one from KDD, and one from MPT. In addition, many of the staff members within each laboratory are part of a particular NTT organization. Thus there is a great deal of vertical information flow, and for similar reasons the laboratories are rather

independent. Practically, though, ATR can be considered as part of the NTT family.

The four ATR laboratories are colocated in a building in Kansai Science City, a new area about halfway between Kyoto and Osaka. I was told that building costs were about $70M. Inside the ATR building there is a fifth laboratory [called NTT Communication Science Laboratories (CSL)], which started operation in July 1991. This operation is not part of ATR but is just leasing the building space. Right behind ATR, NTT owns land, where they may establish a new building for this laboratory. If this happens, it will be the only official NTT laboratory outside of the Kanto area. Almost all the researchers now at CSL were transferred from NTT laboratories in Musashino or Yokosuka in the Tokyo

area.

When I first visited in 1990, ATR was the only building in sight. Now it is surrounded by others in various stages of construction, including several hundred residential townhouses. Across the street a large Sumitomo is going up. Across from that, a big project called the Keihanna Convention Center is under construction. (KEI-HAN-NA is a combination of the abbreviation of three words, Kyoto, Osaka, and Nara.) It will be a convention center with hotel and so on. Those readers who have been to Japan should imagine what Tsukuba Science City (near Tokyo) must have looked like during its earliest days to get a sense of the area around ATR. Like Tsukuba in those days, many

ATR staff members feel isolated from more established parts of Japan, but this will change as the community develops roots.

Almost half the ATR management are on rotational assignment from a couple of NTT's laboratories or divisions. And the other half are mostly from laboratories of KDD (international telephone company), NHK (seminational radio and TV corporation), and CRL (Communication Research Laboratory of MPT). Most of the staff members are also rotational, coming for a few years. About one-fourth are from NTT, but other Japanese companies are also represented. There are also visiting Western researchers. I met four, three Americans and a Frenchman. Their appointments were temporary, 1 year at a time, but apparently could be extended several times by mutual agreement. Each of the four laboratories has a few senior scientists who are either permanent or at least on very long term assignment. These people provide the technical leadership and continuity.

Rotation of staff in and out of the laboratories provides a collection of new ideas but also leads to some underutilization of equipment that is ordered for use in one near term research project but is of less interest to the next. However, even in such cases there is residual benefit, as the overall research projects are generally well planned and are all heading toward some very long term goals, with natural near term revisions and adjustments. I was told that most equipment will be used in succeeding projects. However, staff turnover is definitely ATR's major organizational concern, and I was told that there are new plans to develop more permanent positions. This is a difficult issue, because the ATR laboratories are not permanent. If ATR hires any permanent employee they have to provide for the employee's longterm career plan. I was told that one approach is to hire only those that can

easily be absorbed back into a sponsoring company, a university, or a research institute when the ATR projects end.

Staff experience is also an issue. As is typical of Japanese companies, there are relatively few employees with Ph.D. degrees. Companies that support ATR clearly use the laboratories as a place to put younger employees for training. There was some difference of opinion among the Japanese I spoke to as to whether an assignment at ATR was in anticipation of good work to follow (back at the home company), a reward, or a banishment. This is very company dependent. However, most companies send their good engineers because they are representatives of the company and are representatives of the company and they want them to look good.

Researchers who come with definite ideas about what to work on seem to fare quite well. I was told that each and every researcher knows what he will do at ATR because there are negotiations on this subject before he joins. However, there is the usual trial and error associated with research topics. Perhaps Western researchers have even more flexibility in this matter. I believe, though, that many of the younger Japanese are assigned to projects. Most new staff will work on a project suggested by an advisor or department head, by joining an existing project, or by doing research alone. There is also the problem of what some of these Japanese will do when their ATR assignment is complete. Of course, most will return to their companies, but several have gone on to universities. One expressed concern to me that the basic research he was doing at ATR might not be supported in his home laboratory and hoped that he would be able to find a good spot when he returned. On the other hand, some returning employees from Sony, Matsushita, and others have been known to get overseas assignments (e.g. to the United States, Germany, the United Kingdom, etc.) after their ATR assignments. So

coming to ATR is a kind of stepping stone for them. And, in general, many researchers have been able to make excellent use of the experience, results, and ideas, etc., from ATR and, in some cases, were placed in very good positions. This is especially true for NTT. On balance, a few years at ATR seems to be a good thing for the scientists involved.

The four laboratories are mostly independent, although they are housed in the same building. The independence is enhanced by giving each laboratory its own president as mentioned earlier. This was a bit strange to me, as the total staffing of 260 is low enough that there could be a great deal of interaction, although each of the laboratories is aiming at its own specific research goals. However, given the complicated funding arrangements, it is not too surprising.

A good example of the independence is the use of the Connection Machine. This was purchased primarily for the computational studies of cognitive processes carried out in the Auditory and Visual Perception Research Laboratory. While it is extensively used, I was told that some time is available (there might be some differences of opinion about this), but researchers from other laboratories and even other groups within this laboratory are occasionally discouraged from using it. Dr. Nakane explained to me that funding for the machine is for specific research jobs (rather than for other purposes or mere interest). This requires careful accounting and designation of users. Still, such an important resource ought to be of great help to scientists throughout ATR.

Laboratory staffing is as follows. In 1990, total staff was 262 and total budget was ¥7.5B, about $50M. This does not include the space occupancy fee that ATR International charges to each of the research and development (R&D) organizations. I mentioned this figure to a colleague who had spent 3 years at

[merged small][merged small][merged small][merged small][merged small][merged small][merged small][subsumed][ocr errors][subsumed][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small]
[blocks in formation]

• Measurement of EMG, articulatory movement and acoustics for speech articulator and neural net model.

Computational model of visual cortices and sensor fusion:

• Use of CM-2 to implement Markov random field (MRF) model of images and learning of potential energies for hidden variables without teachers.

Hierarchical MRF based on conditional probability and application to pattern recognition.

• Shape from shading based on forward and inverse models of optics.

• Integration of visual, somatosensory, and motor control information for object manipulation.

There is a great deal of neural net research in progress, including new algorithms and applications to physical situations of speech and vision. The traveling salesman problem is another application being studied. The CM-2 is used heavily now for simulations. (M. Hirayama showed me several very interesting demonstrations.) During my first visit, the ATR staff members were still learning about this machine and usage was low. I was told, however, that the machine is now much busier. It is extensively used as a neural network simulator for the previously mentioned studies of human perception, motor control, and some other related topics. Some of the computations/simulations on CM-2 require an enormous amount of machine use, say, 10 days of continuous time, for example. Presently the CM-2 is working in "exclusive mode" but not in "TSS mode." This is partly because of some problems with the TSS operating system (OS) and partly because of some decrease of computational performance in TSS mode,

which seem to decrease the efficiency of the CM-2 usage. However, the newest version of TSS OS seems to have resolved these problems and is to be installed very soon. However, it has been used heavily enough so that requests are being heard that it be replaced with a more powerful one. The current machine was purchased on a 3-year installment plan and included 32-bit floating point units. ATR is hoping to get one with 64-bit floating point units in the future, but at the current time this has not happened. (I believe that there are three Connection Machines in Japan--one at ATR, a second at the Institute for Computational Fluid Dynamics in Tokyo, and a third at an undisclosed location, probably a private company.)

One of the most interesting aspects of ATR's research is how tightly coupled it is to biological models of real sophistication. A significant aspect of the work seems to be associated with modeling and then verifying models with complicated physical measurements. One example of this is related to understanding the mechanism behind the articulatory to acoustic transform, i.e., to understand the process by which linguistic intentions become speech. Speech production entails extraordinary coordination among diverse neurophysiological and anatomical structures. These unfold through time to produce a complex acoustic signal that conveys to listeners something of the speaker's intentions. ATR's research has used measured movements of the articulator (upper and lower lips and horizontal and vertical jaws) and EMG data from four separate muscle groups and used these to train a neural net to generate motor commands to the articulators.

Another especially elegant example is work by Kawato on studying the trajectory followed by human arms. Kawato proposed that this minimizes

the time integral of the square rate of
change of the torque. Based on this
model he has developed a neural net-
work that reproduced Fitts law (some-
times called speed-accuracy trade-off
of arm movement) and also used it to CONCLUDING REMARKS
study feedback-error-learning for ill-
posed problems. Human sensory-motor
control, such as arm movement, involves
a number of ill-posed problems.

substantial flow of visitors between ATR
and MIT. (Immediately after my visit,
T. Poggio from MIT arrived to give
several days of lectures.)

My visits to ATR have been cordial and revealing. About 10 scientists I met were enthusiastically willing to describe their research, although the independent laboratory structure made it difficult for me to get a full picture. My visits have been highly organized and tightly time constrained. The laboratory I visited is doing very excellent basic research, but I still haven't seen the work on interpreting telephony or communications systems in other laboratories, and I would like to go back and learn about these. What I have seen is a group of capable Japanese, supplemented by a substantial number of long-term Western visitors. There are still some birth pains both at ATR and in the surrounding area; ATR is barely 5 years old. Finally, procedures need to be established to insure that ATR is not only a training ground for young Japanese scientists and a tempo

I think that it is unusual to find so much biophysical expertise being put to use in modeling, especially coupled with advanced computer hardware and software. In the area of vision, there are software. In the area of vision, there are projects to automatically generate 3D images from a series of stereo-pairs and another fascinating project to analyze, quantify, and predict our response to optical illusions. Projects like these require careful eye-movement measurements and distance monitoring, as well as a deep understanding of current ideas in cognition. These are difficult mathematical problems, too. Our brain essentially solves the inverse problem of reconstructing a 3D visible surface from the data of a twodimensional (2D) image projected onto the retina. (Solutions of this problem are related to extensions of regulariza-rary or sabbatical station for Western tion techniques, some of which are very tion techniques, some of which are very well known in computational mathematics.) There certainly is plenty of hardware available for experimenting. While I was there two technicians from Canada were visiting ATR to help set up a pair of Optotracks, sophisticated distance measuring equipment costing in the neighborhood of $60K each. Workstations are everywhere, along with associated peripherals. The ATR building and laboratories are large, spacious, and very well equipped. One of the only places in the United States that I know of, where comparably subtle work is going on, is at the Massachusetts Institute of Technology's (MIT) brain research department, and there is a

ones but also to insure that ATR will develop the continuity and world-class stature that will enable it to accomplish the major long-term goals it has set for itself.

« iepriekšējāTurpināt »