From: hpcc@aquinas.csl.uiuc.edu (Benjamin W. Wah) Subject: Prelim. Report -- NSF HPCC Workshop on Vision, Nat. Lang./Speech Proc., AI Sender: fpst@hubcap.clemson.edu (Steve Stevenson) Date: Fri, 6 Mar 92 13:17:09 CST Approved: parallel@hubcap.clemson.edu This report is posted on multiple electronic boards so it will have wide access to researchers interesed in the HPCCI. B. Wah University of Illinois Distribution: Bulletin Boards: comp.ai comp.ai.neural-nets comp.ai.nlang-know-rep comp.ai.vision comp.arch comp.graphics comp.graphics.research comp.parallel comp.realtime comp.robotics comp.sys.super comp.ai.nlang-know-rep comp.ai.vision comp.graphics.research comp.parallel Mailing lists: IR-L@UCCVMA.BITNET supercomputer@nyu.edu ______________________________________________________________________________ Preliminary Report Workshop on High Performance Computing and Communications for Grand Challenge Applications: Computer Vision, Natural Language and Speech Processing, and Artificial Intelligence 1. INTRODUCTION This article reports preliminary findings of the Workshop on High Per- formance Computing and Communications (HPCC) for Grand Challenge Applica- tions: Computer Vision, Natural Language and Speech Processing, and Artifi- cial Intelligence. Under the support of National Science Foundation, this workshop brought together 23 invited experts from academia and industry. The goal of the workshop was to identify near-term (within five years) and long-term (beyond five years) problems and potential approaches/research directions in supporting grand challenge applications in computer vision, natural language and speech processing, and artificial intelligence (AI) by HPCC systems. Attendees focused on answering the following questions. a) What grand challenge applications in computer vision, natural language and speech processing, and AI can benefit by the availability of HPCC systems? b) How should HPCC systems be designed so that they can support grand challenge applications in these areas? Preparation of the workshop started in late January 1992. Over 40 experts in the three areas and 19 program directors from National Science Foundation were invited. The workshop was held on February 21 and 22, 1992 in Arlington, Virginia, with 23 experts from academia and industry attend- ing and 12 program directors from National Science Foundation serving as observers. Participants in the workshop were divided into three areas, with a vice-chair identified for each. Before the workshop, each vice-chair soli- cited position statements from members of his area, and coordinated the discussions of issues. Separate discussions in the three areas took place on the morning of February 21. In each area, the vice-chair first presented an overview of issues, followed by a short presentation by each ____________________ This workshop was supported by National Science Foundation under grant IRI-9212592. Ideas reported here do not reflect the official position of the sponsoring agency. Preparation of this report was coordinated by Benjamin W. Wah, Thomas Huang, Aravind K. Joshi, and Dan Moldovan. Questions regarding this arti- cle can be directed to them or to any of the attendees listed in Section 3. March 6, 1992 2 member of the area, including the vice-chair. Based on comments received during these presentations and further extensive discussions on the after- noon of February 21, the vice-chair, in consultation with members of the area, prepared a summary report. These reports were presented by the vice-chairs on the morning of February 22 and led to considerable discus- sion. The next section contains a summary of the ideas discussed on Febru- ary 22. The final report, to be released in late April, will be prepared on the basis of this preliminary report and further discussions among the participants through electronic mail. This report contains a collection of ideas expressed by individuals at the workshop; it does not necessarily represent a consensus among all the participants. Further, ideas expressed in this report do not reflect the official position of the sponsoring agency. 2. SUMMARY OF IDEAS 2.1. Computer Vision Area Computer vision has two goals. From the engineering viewpoint, the goal is to build autonomous systems that can perform some tasks that the human visual system can do, and even go beyond the capabilities of the human visual system in multimodality, speed, and reliability. From the scientific viewpoint, the goal is to develop computational theories of vision, and by so doing, gain insights into human visual perception. Grand challenge applications in computer vision fall in two classes. a) Autonomous vision systems have many important applications. Examples include i) flexible manufacturing, ii) intelligent vehicle highway systems, iii) environment monitoring, and iv) visual man-machine interface and model-based compression for telecommunication, multimedia, and education. Note that most of the applications involve interaction of the vision system with the environment and humans. b) Computer vision techniques can also be invaluable tools for studying many basic scientific problems in other areas. A prominent example is the visual understanding of turbulence in fluid flow. The basic scientific issues underlying the applications are i) machine learning, ii) surface reconstruction, inverse optics, and integration, iii) model acquisition, and iv) perception and action. HPCC support for computer vision can be divided into three classes. 1) Vision Systems. There are two cases: i) designing vision systems, and ii) running vision systems. Both require huge amounts of computation power and memory. In addition, vision systems often require real-time operation, low-cost, low power, small volume, and low weight. For instance, a vision system may receive as its input 1-100 gigabits/second of image data that need to be processed in real time. 2) Vision Tasks. Tasks in a vision system fall into roughly three categories: low-level (e.g., noise reduc- tion, data interpolation, feature extraction, and matching), intermediate- level (e.g., grouping), and high-level (e.g., object recognition). To per- form these tasks efficiently, each level may require different types of March 6, 1992 3 computer architectures. Therefore, for many vision systems, a heterogene- ous parallel architecture may be the best answer. Of particular interest is the scalability of such architectures, especially the question of how the different components can be easily ``glued'' together, and the communi- cation and control pathways between the different homogeneous parallel pro- cessors. Another challenge is to develop easy-to-use software for such architectures. 3) Distributed Processing. In many vision systems, compu- tations need to be carried out at several different locations. Thus, dis- tributed computing is of great importance. One aspect of this problem is the transmission and management of huge amounts of image data. Computer vision is related to other grand challenge areas because a) many applications, such as video compression and man-machine interface, involve both vision and speech; and b) AI techniques, such as knowledge- based reasoning, are needed in vision systems. Infrastructure supports for computer vision include a) sharing image databases, software over high-bandwidth networks, and b) providing facili- ties and incentives for architects and computer-vision researchers to work together. 2.2. Natural Language and Speech Processing Area Grand Challenge applications in this area include a) electronic libraries and librarians, which include the use of spoken language inter- faces, machine translation, and full text retrieval, and b) spoken language translation. The fundamental scientific and enabling technologies include a) corpus based natural language processing (NLP) that involves the acquisition of linguistic structure, b) statistical approaches to NLP, c) language analysis and search strategies, d) auditory and vocal-tract modeling, e) integration of multiple levels of speech and language analyses, f) connec- tionist speech and language processing, g) full text retrieval techniques, and h) special-purpose architectures. Bridges to other grand challenge areas include a) optical character reader (OCR), b) handwriting analysis, c) document image analysis, d) multi-media interfaces, and e) integration of multiple knowledge sources. Architectural needs for supporting natural language and speech pro- cessing include a) faster processors with larger memory, b) general purpose supercomputing, c) heterogeneous architectures, such as systems including signal processing and symbolic processing capabilities, d) homogeneous architectures not requiring wide floating point arithmetic, such as those for modeling connectionist architectures, and e) high-bandwidth real-time inputs and outputs. Infrastructure supports include a) shareable text and speech data- bases, b) smart compilers and open parallel systems, c) technical staff for developing sharable tools, and d) access to high-performance computing through high-performance wide-area networks. March 6, 1992 4 2.3. Artificial Intelligence/Computer Architecture This area covers the broad field of AI and the computer architectural support for HPCC AI systems. Some of the grand challenge applications are a) nation-wide job banks, b) electronic library, c) electronic market places, d) large-scale real- time planning and scheduling, e) automation in constructing very large knowledge bases, and f) automation of decision making. For example, an electronic library may involve a diverse collection of text, images, data- bases, and other information scattered around the net in an assortment of formats. Users will need an intelligent librarian program to help guide them through all this information. The librarian will need to communicate with users in natural language and understand something about text stored in the network. The basic research issues and enabling technologies underlying the applications include a) study and design of scalable and verifiable ``trad- itional'' symbolic AI/expert systems, b) construction and utilization of very large knowledge bases, c) development of highly parallel machine learning techniques, d) research on active memories as a means of increas- ing the contribution of knowledge sources in reasoning, e) development and evaluation of marker/value passing techniques, f) application of neural networks to AI, and g) further studies of heuristic search techniques applied to problem solving. Some computer architecture implications are a) increased use of mas- sively parallel processing techniques with a goal of achieving real-time AI processing, b) understanding of the computational requirements of various AI paradigms and how they translate into system requirements in order to either build specialized systems or improve the mapping of AI problems into existing high performance computers, c) understanding of the architecture of systems supporting both numeric and symbolic AI problems, d) development of knowledge base management techniques for implementing efficient multi- level knowledge based systems, e) deciding when it is best to use general- purpose versus specialized accelerators, and f) development of compilers for AI languages on today's supercomputers. Required infrastructure supports include a) access to large fast com- puters by the AI community, b) access to on-line large knowledge bases and corpora, c) sharing systems and research results achieved in large projects by the community, and d) development of computational benchmarks for impor- tant AI paradigms. 3. WORKSHOP ATTENDEES Workshop Chair Benjamin W. Wah University of Illinois, Urbana-Champaign wah@aquinas.csl.uiuc.edu Vision Area Thomas Huang University of Illinois, Urbana-Champaign (Area Vice Chair) huang@uicsl.csl.uiuc.edu March 6, 1992 5 John Aloimonos University of Maryland, College Park yiannis@alv.umd.edu Ruzena K. Bajcsy University of Pennsylvania bajcsy@central.cis.upenn.edu Dana Ballard University of Rochester dana@cs.rochester.edu Charles R. Dyer University of Wisconsin, Madison dyer@cs.wisc.edu Tomaso Poggio Massachusetts Institute of Technology poggio@ai.mit.edu Edward M. Riseman University of Massachusetts, Amherst riseman@cs.umass.edu Steven L. Tanimoto University of Washington tanimoto@cs.washington.edu Natural Language and Speech Processing Area Aravind K. Joshi University of Pennsylvania (Area Vice Chair) joshi@central.cis.upenn.edu Ralph Grishman New York University grishman@nyu.edu Lynette Hirschman Massachusetts Institute of Technology hirschman@goldilocks.lcs.mit.edu Stephen E. Levinson AT&T Bell Laboratories sel@research.att.com Nelson H. Morgan University of California, Berkeley morgan@icsi.berkeley.edu Sergei Nirenburg Carnegie-Mellon University sergei@nl.cs.cmu.edu Craig Stanfill Thinking Machines Corporation craig@think.com Artificial Intelligence and Computer Architecture Area Dan Moldovan University of Southern California (Area Vice Chair) moldovan@gringo.usc.edu Doug DeGroot Texas Instruments degroot@dog.dseg.ti.com Kenneth DeJong George Mason University kdejong@aic.gmu.edu Scott E. Fahlman Carnegie-Mellon University scott.fahlman@cs.cmu.edu Richard E. Korf University of California, Los Angeles korf@cs.ucla.edu Daniel P. Miranker University of Texas, Austin miranker@cs.utexas.edu Salvatore J. Stolfo Columbia University sal@cs.columbia.edu Benjamin W. Wah University of Illinois, Urbana-Champaign wah@aquinas.csl.uiuc.edu National Science Foundation Observers Syed Kamal Abdali Numeric, Symbolic, & Geom. Computations kabdali@nsf.gov Paul G. Chapin Linguistics pchapin@nsf.gov March 6, 1992 6 Su-Shing Chen Knowledge Models and Cognitive Systems schen@nsf.gov Bernard Chern Microelectronic Info. Proc. Systems bchern@nsf.gov Y. T. Chien Info., Robotics, & Intelligent Systems ytchien@nsf.gov John H. Cozzens Circuits and Signal Processing jcozzens@nsf.gov John D. Hestenes Interactive Systems jhestene@nsf.gov Richard Hirsch Supercomputer Center rhirsch,@nsf.gov Howard Moraff Robotics and Machine Intelligence hmoraff@nsf.gov John Lehmann Microelectronic Info. Proc. Systems jlehmann@nsf.gov Pen-Chung Yew Microelectronic Systems Architecture pyew@nsf.gov Zeke Zalcstein Computer Systems Architecture zzalcste@nsf.gov