Lapas attēli
PDF
ePub

NEUROCOMPUTING AT
MITSUBISHI ELECTRIC

This article profiles Mitsubishi Electric and describes the work in its

Neurocomputing and Optoelectronics Group in Osaka.

by David K. Kahaner

• Manufacturing Development Labo

ratory (Osaka)--Development of technologies for all of MELCO, including in-company manufacturing equipment and systems, processing, and insulation.

• Materials and Electronic Devices

Laboratory (Osaka)--Development of metals, ceramics, high-density materials, and electronic devices and analysis of existing technologies.

• Industrial Electronics and Systems

Laboratory (Osaka) --Development of factory automation, public utility, and automotive systems and equipment.

INTRODUCTION

• Information, telecommunication,

electronic devices, and systems Western scientists, especially those in universities, often have only vague • Industrial and automotive equipideas about the real size of some Japanese ment high tech companies. A good example is Mitsubishi. While probably best • Heavy machinery (locomotives, known in the West for its automobiles nuclear power equipment, etc.) and shipbuilding activities, I wonder how many realize that Mitsubishi Elec- with the first two nearly equal and tric (MELCO), which does not include accounting for about 60% of sales. either of the automotive or ship con- Research and development (R&D) struction companies, is itself a $27B expenses are as follows in billions of company at today's exchange rate, with yen, approximate dollars (¥125/$1), and just under 100,000 employees, and cele

percent of sales. brated its 70th anniversary in 1991? See my report, “Flexible Automation,"

% of Scientific Information Bulletin 16(1), Year ұВ

$B Sales 27-35 (1991), for a description of some of Mitsubishi's shipbuilding activities. 1987 94.6 1.28 4.8 For a more general overview see 1988 101.9 1.16 4.9 "Mitsubishi Electric Corporation: Its 1989 118.5 0.95 4.4 Basic Research Focus," Scientific 1990 145.1 0.82 4.3 Information Bulletin 16(4), 39-42 (1991). 1991 159.8 0.76 4.9 Note that in January 1992 the Yamato I, a 280-ton prototype superconducting

The main R&D activities continue propulsion system ship, was launched to be electronics, new energy sources, by Mitsubishi Heavy Industries (separate new materials, and biotechnology. company from MELCO), and that MELCO has quite a number of research MELCO and other parts of the laboratories. The oldest, the Central Mitsubishi family will participate on a Research Laboratory (CRL), near bid for the 409-km/100-minute Seoul- Osaka, has been in existence under this Pusan, Korea, bullet train project (total name since 1944. cost is estimated near $10B), to be completed before the end of the century. • Central Research Laboratory MELCO products fall into four areas. (Osaka)--Basic R&D of future

products and of fundamental tech• Consumer products (audio-visual, nologies for existing and new product

home electronics, and appliances) lines.

• Consumer Electronics Laboratory

(Kyoto)--Development of new electronic product lines and mass production technologies for these products.

• Consumer Products Laboratory

(Kamakura)--Development of personal information-communications equipment, functions for home electrical appliances, systemization technologies, new-function electronic products, and productevaluation technologies.

• Industrial Design Center

(Kamakura)--Industrial design activities related to all MELCO products. • Computer and Information Systems MELCO is heavily into advanced trans- and similar hazards. It can be initiated

Laboratory (Kamakura)--R&D of portation, such as superconducting by a phone call. computer- and information-related trains (mag-lev) and superconducting My visit to MELCO was organized systems and equipment.

ship propulsion systems, and it sup- by

plies vast numbers of locomotives and • Communication Systems Labora- related rail-stock to countries from Dr. Wolfgang Banzhaf tory (Kamakura)--Overall R&D of China to Mexico.

Mitsubishi Electric communication systems technolo- MELCO is active in the Institute Central Research Laboratory (CRL) gies and devices.

for New Generation Computer Tech- Neurocomputing Group

nology (ICOT) program with the devel- Quantum Electronics Dept • Electro-Optics and Microwave Sys- opment of several versions of PSI, a 1-1 Tsukaguchi-honmachi 8-chome

tems Laboratory (Kamakura)-- parallel sequential inference machine, Amagasaki, Hyogo 661, Japan Development of new technologies including the first one ready in Japan. Tel: +81-6-497-7050 for components in fields such as One application is for a Japanese- Fax: +81-6-497-7288 electro-optics and

and microwave English translation system. There are a E-mail: ban@adm.crl.melco.co.jp systems.

number of other parallel processing

projects including a high speed data- I had met Banzhaf in the fall of 1991 • LSI Laboratory (Osaka)--R&D of base machine and a parallel-syntax when he presented a proposal at the

design and production technologies processor for man-machine speech New Information Processing Technolfor integrated circuits (ICs), large recognition system. The company has ogy (NIPT) Workshop in Yokohama scale integration (LSI), very large produced more than 150,000 elevators (see "First New Information Processing scale integration (VLSI), and dis- and escalators, including a group that Technology Workshop '91,” Scientific crete semiconductors as well as the uses artificial intelligence (AI), expert Information Bulletin 17(1), 51-60 (1992)). creation of new products.

systems, and fuzzy theory, and it has Dr. Banzhaf is a visiting researcher on

recently produced the world's first spiral leave from the Institute for Theoretical • Optoelectronic and Microwave escalator (installed in San Francisco). Physics and Synergetics (Stuttgart) and

Devices Laboratory (Osaka)--R&D Similarly, it has a fuzzy controller for has been at Mitsubishi’s laboratory since of design and production technolo- electric discharge machines, an expert September 1989. He is working in the gies for optoelectronic and micro- system for insulation diagnosis, and a neurocomputing group, run by wave devices and the creation of developing knowledge media station. new products.

MELCO has the largest value added Dr. Kazuo Kyuma

network in Japan (packet-switching) Manager, Neurocomputing Group • ASIC Design Engineering Center and is planning to use this to link Dept of Solid State Quantum

(Osaka)--Development of custom 20,000 terminals in its corporate E-mail Electronics
LSI design technologies and sup- net, both Unix workstations in techni- (address same as above)
port for the design of LSIs for use in cal divisions and PCs in administrative Tel: +81-6-497-7078
in-company systems.

divisions. A server will be installed in Fax: +81-6-497-7288

a local area network (LAN) used by E-mail: kyuma@qua.crl.melco.co.jp • MELCO Research Laboratory each laboratory and a main server will U.S.A. (Boston)

be installed in the company's main office. Recently, Kyuma has received the

When completed this year, it will connect Sakurai-Prize, the most renown in Japan MELCO is currently constructinga 63 offices and 45 sales companies, for optical sciences, partially for his synchrotron-radiation facility. It has an hopefully reducing paperwork into the work on the optical neural chip (below). active superconductivity program and main office from its current 3 tons each Coincidentally, another German, is working on 64-Mbit DRAMs, a GaAs day.

Eberhard Lange (tel: +81-6-497-7050, semiconductor laser and field effect The company is also developing a E-mail: el@adm.crl.melco.co.jp), is also transistor (FET), an optical neurochip, number of products that will depend visiting this group. As is typical of visita process using ultrafine ice particles on using the ISDN (Integrated Ser- ing Westerners in Japanese companies, as a cleaning agent in electronic device vices Digital Network) now being these appointments are temporary but manufacture, semiconductor ceramic installed in Japan. For example, one can often be continued indefinitely by fiber (30-50 microns), and a very high home automation system monitors a mutual agreement. Kyuma's group has integration digital signal processor. home for gas leaks, fire, unlocked doors,

a total of 18 staff. Of these, five have the latter did not work out and my visit A major ingredient of the Mitsubishi Ph.D. degrees and three more have was restricted to optical computing and approach has been the use of quantized some prospect of obtaining this degree neural nets.

learning models. By that we mean that in the future. Typically, Japanese com- The most exciting project here has neural connection weights are not panies hire recent graduates with the been the development of a fully optical permitted to take on continuous values equivalent of a bachelor's or master's neural chip, with eight neurons and but instead are restricted to a finite set, degree. In the case of Kyuma's group, (optically) adjustable weights. Dr. often just two or three values. A quanstaff members have backgrounds in Kyuma has been the key person on this tized learning rule for back propagaphysics, applied physics, mathematics, project, although I was shown the device tion is as follows: materials, etc. There is very little recruit- by Mr. Jun Ohta (tel: +81-6-497-7049, ment of university trained Ph.D. holders. E-mail: ohta@qua.crl.melco.co.jp). (1) Start with random weights (conOn the other hand, scientists working (Ohta is a good example of how a

tinuous). at an industrial laboratory have the company like MELCO increases staff opportunity to obtain a Ph.D. after a skills. Next fall, Ohta will move to (2) Quantize into several discrete levels. number of years and an appropriate Boulder, Colorado, to join Professor amount of published research. In the Thomas Cathey's optoelectronics group (3) Set spatial light modulator to value United States it is also possible to obtain as a postdoctoral student.) The of value of discrete (hardware) a Ph.D. while working, i.e., off campus, Mitsubishi chip is the most advanced weights. but these (U.S.) students are usually to my knowledge. It was developed using well known to their faculty advisor and molecular beam epitaxy (MBE) crystal (4) Present training and supervised there is a great deal of interaction growth technology and GaAs optoelec- signals to network. between student and university. In Japan, tronic device technology. The first the student-university connection is prototype, about a 1-foot cube, was (5) Calculate error by conventional much weaker.

developed in 1988. The current chip is back propagation method. I've often wondered how it is possi- mounted on a board together with other ble for Japanese companies to produce LSI devices and is in use in a demon- (6) Correct continuous weight by so much interesting science with this stration neural computer hooked to a adding error. kind of a system. One part of the reason small workstation. The chip consists of is that good companies are extremely three stacked layers: a light emitting (7) Repeat (2)-(6) for all training sigselective and only hire a few students diode (LED) array, in interconnection nals until connections converge. each year-hopefully the very best. These matrix, and a photodiode (PD) array. then stay with their group for a long Both LED and PD arrays are placed in In this approach, the quantized weights time, gaining experience, attending a crossbar with a dynamic spatial light are built on the optical hardware and meetings, etc. Promising staff are often modulator sandwiched between them the continuous weights are stored in sent outside the company for training as the interconnection weight matrix. memory, allowing continuous changes or additional education (see below). The original neural network used spa- in the weights and parallel optical The group and project leaders are more tial light modulators. The new neuro- implementation at the same time. senior and these people usually have chip uses variable sensitivity of its Mitsubishi staff have shown that learnadvanced degrees, although often photodiodes, controlled externally, to ing can be successful even if synaptic obtained via industrial publications as adjust the weights. This allows vector

adjust the weights. This allows vector weights are quantized into only two mentioned above. (Sometimes indus- matrix multiplication to be done in levels. They have not only many simutrial scientists leave their company for parallel, and because the weight matrix lation results but also implementation academic positions.) The good Japanese can be adjusted, it also means that the experiments using binary operating research laboratories are also sprinkled chip can learn. This is quite distinct semiconductor light modulators (SLMs) with Western visitors. To be perfectly from other approaches in which the to recognize the 26 alphabetic charachonest, the equation still does not interconnection weights are fixed in ters (Ref 1 and 2). compute for me, but the results speak hardware and use Hopfield models Several of the group members for themselves.

without learning ability. The current commented that they appreciated the Our original plan was to learn about chip is an 8x8. Ohta claims that their opportunity to spend time talking to MELCO's work in both neural nets construction approach allows for about me and also listening over my shoulder and software development. However, 2,000 neurons/cm2.

while their colleagues described their own research work. In fact, they felt Another problem with some con- on workstations, sometimes used as a that such interactions were very uncom- nection to optical computation was cluster. Given the large amount of mon. This was extremely surprising to shown to me by Lange. He has been simulation and the sophistication of me, as is typically Japanese, the group studying the issue of identifying the the models, I am surprised that they works in a large airy room with modest face value of a postage stamp from its don't have access to a parallel machine, partitions between desks, and I expected image. The interesting wrinkle here is such as a CM-2. The Industrial Electhat very free interchange would be the that the stamp image is not digitized. tronics and Systems Laboratory does rule rather than the exception. I was Rather, a small number (32) of sensors have an NCube, but the neural net told that there is plenty of joint work directly read intensity and frequency group does not make use of it. but very little free time for researchers from parts of the stamp, and it is their who are not working directly together values that are input to a neural net, REFERENCES on a project to sit around and talk whose ultimate output is the stamp's about their work informally. An inter- face value. The stamp recognition project 1. J. Ohta, Y. Nitta, and K Kyuma, nal seminar was started last year but has succeeded (at least partially) in “Dynamic optical neurochip using hasn't succeeded too well. Work hours optical implementation. The neural variable-sensitivity photodiodes,” Optics are for working. Lunch is normally at network (competitive) has been imple. Letters 16(10), 744-46 (May 1991). one's desk, and seminars are after 1700, mented. Only the sensor device has not there is a commons room, but it is (so far). There is an internal report on 2. M. Takahashi, M. Oita, S. Tai, apparently rarely used. I remarked that this project, which is going to be pub. K. Kojima, and K. Kyuma, “A quanI didn't find it troublesome to have lished as part of a book (Ref 4). tized back propagation learning rule seminars late in the day and that to a Other work has been directed toward and its application to optical neural large extent lack of time to talk to studying time varying spatial patterns, networks,Optical Computing and colleagues was self-imposed. While there and Banzhaf has developed a network Processing 1(2), 175-82 (1991). was general agreement, the scientists capable of removing distortions of also felt that seminar times were sym- patterns in time (Ref 5).

3. W. Banzhaf, M. Takahashi, J. Ohta, bolic, and the atmosphere discouraged There is also work on genetic, evo- and K. Kyuma, “Weight quantization the kind of cross-group interaction that lutionary, or molecular algorithms in Boltzmann machines,Neural is typical at Western research labora- applied, for example, to the traveling Networks 4, 405-9 (1991). tories or at universities.

salesman problem (TSP), where the The neural net group is mostly doing shortest tour in a given distribution of 4. W. Banzhaf, E. Lange, M. Oita, algorithm development and simulation cities is looked for which visits all cities J. Ohta, K. Kyuma, and T. Nakayama, of different learning models. Some of once and returns to the starting point. “Optical implementation of a competthis is clearly motivated by the parallel This problem is "classical,” but it is a itive network,” MELCO Internal work on the optical neural chip described prototype of many nonpolynomial (NP) Report, to be published in Fast Intelliabove (Ref 3). The focus of this work complete (i.e., problems that cannot gent Systems, B. Soucek, editor (Wiley has been to try and solve the perfor- besolved in time, which is a polynomial Series in Sixth Generation Computing mance degradation problem that is in the number of cities).

Technologies). usually associated with weight quan- The group has access to MELCO's tization. (Performance here is expressed Cray Y-MP, which is in the Materials 5. W. Banzhaf and K. Kyuma, “The by the number of steps necessary for

and Electronic Devices Laboratory, but time-into-intensity-mapping network,” successful training.)

most of their simulation work is done Biological Cybernetics 66, 115-21 (1991).

« iepriekšējāTurpināt »