Lapas attēli
PDF
ePub

The characteristics of the NIPT Program: information in the actual world, and The workshop began with a half day

passes appropriate and advanced judg- of plenary talks in which the four sub(1) The massively parallel system, ment and problem-solving.” Some of group moderators gave their views of

neural system, and the integrated the documentation provided at the the eventual research thrusts in their system are constructed and are workshop also describes impressive areas, followed by 2 days of four paralcomposed of millions of element goals. However, based on the submitted lel sessions, one on each of the four processors. Robust system com- workshop proposals, the fog is slowly subgroup topics. On the last day a general posing technology and the adap- lifting as scientists begin to specify the session gave the moderators an opportive control technology for load explicit research that they hope to tunity to present a very rough master distribution to the element pro- perform. It is my opinion that what is plan for each group. These were done cessors are needed.

likely to emerge is quite reasonable, under a great deal of time pressure (the

and that in many ways the program is session moderators are to be congratu(2) Important subjects are to develop conservative.

lated for achieving as much as they the massively parallel language and

did), but these plans had not been the massively parallel operating DISCUSSION

coordinated and I felt that they needed system. In the languages the describ

much further refinement. ing ability is requested to enable In this workshop proposals were I should mention that although the direct mapping of the problem broken into one of four subgroups, referred to as proposals, in fact most of on the system and to enable the Theory and Novel Functions for Appli- the presentations were more like autonomous cooperative actions cations, Massively Parallel Systems, extended abstracts, without budgets, of each element processor. In the Neural Systems, and Optical Comput- specific task descriptions, milestones, operating system, the learning ing Systems (corresponding to T,MN, evaluation criteria, etc. Individual function is requested to control and O, respectively). I believe that this proposal descriptions were labelled as the whole system in a robust and stems from NIPT views that show opti- not for distribution. However, proautonomous manner.

cal technology as a fundamental tech- posal titles and authors were not so

nology underpinning both a massively labelled and these are in the Appendix. (3) In order to construct the system parallel and (perhaps equivalent) neural Interested readers are encouraged to

composed of millions of element system which, in turn, uses new theory contact the authors directly for more processors, consideration should

to help develop new and important detailed information. be made to the technology of inte- applications. Earlier documents had My understanding is that the steergrated circuits to carry plural ele- listed eight basic areas with some over- ing committee will now attempt to decide ment processors on the wafer and lap; at the moment these have been re- which of these are to be moved forthe application of the optical device aggregated into four. With the help of ward, and these individuals will be asked technology with ability of massively the authors, the organizers tried to place to produce more detailed proposals. I parallel transmission and process- the proposals fairly, but several placed think there will be many changes resulting of information.

in T could have as easily been put into ing from the current workshop. Some

N. Nevertheless, at this stage it is fair to changes were announced even before In my earlier reports I have com- say that there is very little integration the workshop was ended. Original plans mented repeatedly on the vagueness of between the four sections, and this will were to have two additional feasibility the proposed NIPT (RWC, 4DC) pro- be a major task for the steering com- workshops, in January and March 1992. gram. Some of this vagueness may have mittee. (This opinion was quite gener- It is now the intention to have one resulted in a rather imposing view of ally shared by the participants; some more workshop during the week of what might be accomplished. For exam- even wrote as much in their proposals.) 2 March, at which time an integrated ple, in the NTW mentioned above, it is Perhaps it will be impossible to do master plan will be presented. Project written that “the Japanese are looking more than give lip service to this inte funding will begin the following month. for ways to make a million processor, gration, as the time constants for some Most of the decision making will go on ultra-supercomputer, artificially intel- of the research proposals were vastly in executive sessions, and the general ligent thinking machine." MITI's own different. Also, as yet there has been no feeling among the attendees was that documentation hopes that this program attempt to integrate the proposals within individuals with whom collaborations will be “a means of realizing real-world a subgroup; certainly the Western sci- were to be established had already mostly computing that flexibly deals with entists had no idea of what they would been selected. There will certainly not imperfect, ambiguous, and changing hear when they arrived.

be a refereeing process in the Western

etc.

sense; however, the committee mem- connections have not been so well estab- readiness. During the final plenary bers are very senior scientists and readers lished. More senior faculty, he claimed, session, the draft plan actually did not should expect that the projects funded have many opportunities for support explicitly highlight bias towards a datawill be internationally credible. It is from government and industrial sources. flow machine but listed dataflow as not entirely clear, though, how new At least one European observer felt one option in a list of options that ideas that surface during the 10-year that the Ministry of Education would needs more study. duration of the program will be handled. not stand by while MITI infringed on Expectations of the program, accordI mentioned above that participating its jurisdiction of university funding. ing to the steering committee, are still in this project is a high prestige item. The most concrete of the sessions focused on integration of (imperfect) This was echoed by a German colleague concerned massive parallelism. Herel information, adaptation, learning, who felt that the financial support, was told repeatedly by attendees that approximate optimization, etc. Specific while welcome, was really incidental. the focus is that ETL wants to continue functions that were mentioned in the It is presumably the reason that the its research on generalizations of the overall plan include speech understandKorean proposal in the Theory subgroup dataflow model that it has been work- ing based on constraint satisfaction, was presented with a clear statement ing on for some time and for which the flexible action control in robots, as well that only technical collaboration, not laboratory has a great deal of hardware as applications to massively parallel, financing, was being requested. In the experience. They want to build a 10,000 integrated, neural, and optical systems, Optical Computing Systems subgroup, processor 1-TFLOP machine within another Korean proposal on Optical 5 years as a prototype and a one million The theory session was composed Neural Systems seemed to indicate an processor machine within 10 years. As of about 20 essentially independent interest in obtaining advanced Japanese there are already commercial machines proposals, ranging from a 50-year prooptical devices as part of the interna- with tens of thousands of processors posal from Mitsubishi research for tional collaboration. Similarly, a (albeit simple ones), this does not appear computer vision understanding; Singaporean was seeking funding but to be a revolutionary step, although R. Eckmiller's proposal to deal with probably would be happy just to collab- there are many, many problems to be international issues such as smog orate technically. There is also good studied concerning the computing modeling, global warming, etc.; down reason to think that prestige, as well as model, reliability, and software devel- to a 3-year proposal from Singapore's being “plugged in,” is a major factor in opment. The emphasis in this develop- Institute of Systems Science (ISS) for most Japanese company participation. ment will be a machine that can adapt development of what will essentially

MITI has made significant efforts to various required computing para- be a product. Most of these papers to make this a program that is not only digms and so will function as a general represented good science, but even if focused on Japanese industry. In July purpose resource. Nevertheless, the all are funded and all were to succeed 1991 [“NIPT Feasibility Study and working language will probably be some nearly to their proposers' expectations, Workshops,” Scientific Information type of parallel extended object-oriented the fundamental goals of the program Bulletin 16(3), 5-6 (1991)] I reported language. I was told that at least one may not be attained. Thus readers should that there was a major change in the Japanese company proposal would likely not expect that whatever hardware and way ownership of intellectual property be withdrawn (although they would software are eventually built within the was to be handled, with equal sharing probably continue as an industrial scope of this program will result in a between the Japanese Government and partner), as it seemed clear to the pro- system that is able to realize when we the inventor. It is also important to posers that ETL had its own view of are telling a joke. Further, it does not note that for the NIPT project there is how it wanted to proceed. After listen- seem that the most advanced, hence now the possibility that Japanese uni- ing to the individual proposals, the ambitious, of the research from this versity researchers can obtain funding session moderator, Dr. T. Shimada of subgroup could have much direct impact from MITI. This could supplement ETL, presented a draft master plan (of on the design of the massively parallel funding from the (much poorer) Ministry his session) to the entire workshop. system, probably coming too late in the of Education. One Japanese professor This draft does not suggest that optics cycle. told me that this possibility of new will play a major role, except perhaps at I asked about computing resources funding was the major reason he was the level of interconnects. This is sensi- that would be made available to collabparticipating. Another told me that this ble, as the hope is to begin building in orators in this project. (The a-building source of new Japanese Government the next fiscal year, whereas optical massively parallel system will not be funding for universities would be of technology, while very promising, does ready at the early stages as a research most benefit to younger faculty whose not seem to be at the same stage of tool for users.) The organizers have not settled this question but spoke of making commercial parallel systems available and also of providing a high speed network for communication and experimentation among the participants.

The neural session resulted in a draft master plan stating that they hoped to build a 10-TCUPS machine in 10 years. Last spring, Hitachi demonstrated a working 2-GCUPS wafer scale neural system. Thus 10 TCUPS appears to be a realistic target within 10 years, and certainly not an unexpected one. One change coming out of the workshop was that this committee has decided not to validate the system on real applications but only on "novel functions for applications,” i.e., application building tools.

The optical computing systems session emphasized the optical interconnect technologies and the supportive device and materials processing. The sentiment was that this type of work needed to be done before realistic sizings of processor architectures could be made. Of the 13 Japanese proposals, 8 were really expansions of and capitalization on existing industry (6) and university (2) optoelectronic device R&D. There was a ground swell of opinion that the session should report back only on devices for interconnects and leave the other two assigned topics (optical digital and optical neural) for other subgroups (namely, massively parallel and neural systems). The session moderator agreed with the technical reasoning, but said it would have to be discussed after the workshop. Virtually all the major Japanese companies, except NTT, were represented with these types of optical device proposals.

Appendix

PROPOSALS PRESENTED AT THE FIRST NIPT WORKSHOP 1991

Session T1 Theory and Models (1)

T1-1 Computational Learning Theory of Probabilistic Knowledge Representations

Abe, Naoki (NEC Corp., Japan) (ABE@IBL.CL.NEC.CO.JP)

T1-2 Ecological and Evolutional Computation Models for Massively Parallel/Distributed Systems

Matsuo, Kazuhiro (Fujitsu Laboratories Ltd., Japan) (MATSUO@IIAS.FLAB.FUJITSU.CO.JP)

T1-3 Pattern Recognition Using Neural Associative Memory

Lui, Ho Chung (National Univ. of Singapore, Singapore) (ISSLHC@NUSVM.BITNET)

Session T2 Theory and Models (2)

T2-1 Fundamental Research on Information Integration Technology Based on Multi-Neural Network

Matsuba, Ikuo (Hitachi Ltd., Japan) (MATSUBA@SDL.HITACHI.CO.JP)

T2-2 A Vision Processor in Neural Architecture

Banzhaf, Wolfgang (Mitsubishi Electric Corp., Japan) (BANQRELACHE.ADM.CRL.MELCO.CO.JP)

T2-3 Hyper-Heterarchical System for Auditory Scene Understanding

Kawahara, Hideki (Nippon Telegraph and Telephone Corp., Japan) (KAWAHARA@SIVANTT.JP)

Session T3 Recognition and Understanding

T3-1 Sensory Integration

Sugie, Noboru (Nagoya Univ., Japan) (SUGIESUGIE.NUEE.NAGOYA-U.AC.JP)

T3-2 Research on Information Fusion Methods in Pattern Recognition/Understanding

Tsukumo, Jun (NEC Corp., Japan) (TSUKUM@PAT.CL.NEC.CO.JP)

T3-3 A Multi-Modal Human Interface

Chiba, Shigeru (Sharp Corp., Japan) (SCHIBA@TOKEN.SHARP.CO.JP)

T3-4 Action Image Database

Abe, Hasahiro (Hitachi Ltd., Japan) (ABE@CRL.HITACHI.CO.JP)

T3-5 Recognition of 3000 Chinese Characters Handwritten Kanji Characters

Muehlenbein, Heinz (GMD, Germany) (MUEHLEN@GMDZI.GMD.DE)

Session T4 Multimedia Information-Base and Interface

T4-1 Self-Organizing Information Base

Hirayama, Masaharu (Mitsubishi Electric Corp., Japan) (HIRAYAMA@SYS.CRL.MELCO.CO.JP)

T4-2 Multilingual Natural Language Interface/Interaction for Human and Computer

Ahn, Byung-Sung (Electronics and Telecommunications Research Institute, Korea) (D6000@KIET.ETRLREKR) T4-3 Some Novel Functions for the NIPT Computer Interface

Inoue, Hirochika; Sato, Tononasa; Hirose, Michitaka (Univ. of Tokyo, Japan) (INOUE@JSK.T.U-TOKYO.AC.JP)

T4-4 Self-Organizing Adaptive Mechanisms Based on Uncertain Information

Taura, Toshiharu (Nippon Steel Corp., Japan) (TAURA@ELELAB.NSC.CO.JP)

Session TS Inference and Problem Solving

T5-1 Learning Neural Computers for Global Forecasting and Control - Theory and Applications

Eckmiller, Rolf (Univ. of Dusseldorf, Germany) (ECKMILLE@DDORUD81.BITNET)

T5-2 Research on the Dynamic Inductive Inference

Sakurai, Akito (Hitachi Ltd., Japan) (SAKURAI@HARL.HITACHI.CO.JP)

T5-3 Coordinative Distributed Problem Solving Model by Structured Agents

Fukumura, Satoshi (Kawasaki Steel Corp., Japan) (FUKUMURA@SYSTEM.KAWASAKI-STEEL.CO.JP)

Session T6 Integrated Systems

T6-1 Development of a Flexible Hand-Eye Robot in Neural Architecture (refer to the proposal of T3-5)

Muehlenbein, Heinz (GMD, Germany) (MUEHLEN@GMDZI.GMD.DE)

T6-2 Human Interface Using Artificial Reality

Kurihara, Tsuneya (Hitachi Ltd., Japan) (KURIHARA@CRL.HITACHI.CO.JP)

T6-3 Research on Flexible Information Processing Models - An Autonomous Growth System

Asakawa, Kazoo (Fujitsu Laboratories Ltd., Japan) (ASAKAWA@FLAB.FUJITSU.CO.JP)

Session M1 Architecture (1)

M1-1 Research and Development on Composite Super-Fine-Grain Architecture for Massively Parallel Computing

Systems

Hiraki, Kei (Univ. of Tokyo, Japan) (HIRAKIQIS.S.U-TOKYO.AC.JP)

M1-2 A Dataflow Massively Parallel Computer

Shimizu, Uasahisa (Sanyo Electric Co. Ltd., Japan) (SHIMIZU@EDDEN.HIRAKATA.SANYO.CO.JP)

M1-3 Research and Development of a Massively Parallel Architecture Based on the Extended Dataflow Model

Yamaguchi, Yoshinori (ETL, Japan) (YAMAGUTIQETL.GO.JP)

Session M2 Architecture (2)

M2-1 Adaptive Massively Parallel Machines

Koike, Nobuhiko (NEC Corp., Japan) (KOIKE@CSL.CL.NEC.CO.JP)

M2-2 A Massively Parallel Machine with Optical Interconnection

Oyanagi, Shigeru (Toshiba Corp., Japan) (OYANAGI@ISL.RDC.TOSHIBA.CO.JP)

M2-3 Massively Parallel Image Processing Sub-System

Mizoguchi, Masanori (NEC Corp., Japan) (MIZO@PAT.CL.NEC.CO.JP)

« iepriekšējāTurpināt »