Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
- 11 - SYSTEMS, COMMUNICATION, AND SIGNAL PROCESSING The importance of secure, high-speed communications, signal pro- cessing, and image processing capabilities to the Navy, and in fact all the Services, hardly needs elaboration. Recent news items, for ex- ample, about how increasingly quiet Soviet submarines have become, make amply clear the significance of rapid and sophisticated methods for the extraction of information from very weak signals in uncertain and non- stationary environments, and its rapid dissemination to a wide network of users. Moreover, the continued success of the ONR's longstanding and notable efforts in the solid state area make the needs for corres- ponding signal processing efforts even more acute. Without prior attention to the appropriate application, the potential capabilities of very-high-density submicron devices may never be usable. The ONR's electronics program in the area of systems, communi- cations, and signal processing is an important, but still small, fraction of its core research budget. Although the area has recently had a substantial infusion of SDIO/IST funds, and although it is also partly covered by efforts in the mathematics and computer science programs, it is so relevant to Navy needs that additional funding of the core program base would appear to be extremely important. The high quality of the small program currently existing in this area must be commended, but efforts to bring more researchers into the area without reducing the exploding opportunities in solid state electronics will be essential for significant progress. Of the several directions for systems, communications, and signal processing that could be recommended, perhaps the most exciting are those at the boundaries of several disciplines. We recommend algo- rithms for distributed architectures and networks as a major topic that is both forward-looking and at the same time related naturally to the existing efforts in the ONR. This area is divided into four subtopics: o Algorithms and architectures. o Sensor array processing and high-resolution spectral ana-lysis. o Distributed algorithms for network control. o Neural nets. These are described in greater detail in this section, but first some general remarks may be useful. There are many fundamental issues concerning the design of algo- rithms for signal processing, sensor processing, pattern recognition, control, and other functions when those algorithms must operate in a distributed environment. In some examples, such as neural networks and parallel processing, one uses distributed signal processing elements to enhance the computational power and speed of the system, whereas in other examples, such as network control and distributed sensor systems, the geographical separation of the nodes forces the system to be distributed. In both types of examples, however, the design of the communication between the processing elements is at least as important as the processing done by the individual elements. In both types of
- 12 - examples, the distributed nature of the system both permits the opportunity for robust operation in the presence of failures and also provides the challenge of understanding how to construct distributed algorithms that can adapt to failures. Finally, there is also the common problem of how to scale such algorithms with the number of nodes or processing elements. These general questions appear implicity in all the four promising research areas discussed in the following sections. ALGORITHMS AND ARCHITECTURES A look at current activity in the very large scale integration (VLSI) area will show a substantial effort in more sophisticated digi- tal signal processing (DSP) systems. However, most of this activity is microprocessor-based, albeit on ever-newer and more powerful units, and therefore subject to certain fundamental limits on the maximum process- ing speed. For example, microprocessor-based systems for decoding con- volutional codes are currently in the 10-20 Mbits/sec range, and more than doubling this rate appears to be difficult to achieve, whereas applications demand rates of several Gbits/sec. New computing struc- tures are clearly necessary. One possibility is the use of general pur- pose parallel computing machines such as the hypercube, the connection machine, the BBN butterfly processor, the SAXPY-lM, and the GE WARP systolic processors. However, if we look a few years (5 to 10) ahead, it is conceivable that advances in technology and in computer-aided design (CAD) will allow us to build cheaper and faster special purpose solutions for different classes of signal processing applications. The warm reception accorded ZORAN's very simple move in this direction through the introduction of dedicated vector multipliers and adders is an indication of how welcome more customized solutions could be. More evidence accumulates every day (see the August 6, 1987, issue of Electronics). There has been considerable progress already in the direction of special purpose arrays with work on systolic arrays, wavefront pro- cessors, regular iterative arrays, and the like, and, in fact, some of this work has been sponsored by the Contract Research Department (CRD) Electronics Division already (under the USER-VLSI program) and by the CRD Computer Science and Mathematics Divisions. However, these efforts are winding down, perhaps just at the time when increased effort toward longer-term issues behind the original support is most needed. It will become important to move toward the development of a theoretical base for the design of classes of special purpose computing arrays as opposed to the rather ad hoc methods used so far for the invention of systolic arrays (and some generalizations thereof). Bridges also need to be developed to related work in computer science on compiler theory and on the development of methodologies for high- level programming languages, both of which have some surprising (at least initially) similarities to problems encountered in the studies motivated by signal processing. Moreover, such research has to be closely linked with the study of methods of fault modeling, test
13 - generation, error detection and correction, fault recovery, and fault- tolerant design. SENSOR ARRAY PROCESSING AND HIGH-RESOLUTION SPECTRAL ANALYSIS Fast and accurate algorithms for detecting and resolving narrow band signals in noise, while also determining their angles of arrival, are clearly of great interest to the Navy. Although the current ONR program has a few excellent investigators in this field, there is an opportunity here for bringing in several other researchers from related areas in signal processing, numerical analysis, and processor array design, and for useful collaborative efforts with researchers at sev- eral naval laboratories, especially Naval Ocean Systems Center (NOSC), Naval Undersea Systems Center (NUSC), Naval Research Laboratory (NRL), and Naval Ocean Research and Development Activity (NORDA). There has been striking progress in recent years in direction of arrival (DOA) algorithms such as MUSIC, root-MUSIC, and recently ESPRIT. Some useful solutions to the longstanding problems of combat- ing coherent interference (as in multipath or smart jamming) have been developed. However, there is a need for further examination of sensi- tivities to model uncertainties, robustness in numerical calculations, and the development of special purpose architectures of various types (e.g., bus-connected processors with local memory, systolic arrays, or connection machines). Interaction with naval laboratories will also enable some trials with real data, an important guide in such highly theoretical areas. One might mention that such tests have recently been made with ESPRIT on the analysis of very flexible robot arms with encouraging results. There are still several open questions in the area of spectral analysis, which is not surprising considering the very wide range of problems and applications in this field. DISTRIBUTED ALGORITHMS FOR NETWORK CONTROL Communication networks, whether for data, voice, or both, inher- ently require distributed control. The required distributed control algorithms range from link access protocols to routing and flow control algorithms to transport and higher layer protocols. Research and devel- opment on conventional networks has been well funded by the government and by industry over the past 20 years and has led to a mature and recognized research field. This research, and the corresponding tech- nology, is based on a clear separation between communication theory (and technology) and network theory (and technology). The Navy (and more generally the Department of Defense) has a critical need for survivable communication networks, that is, networks that maintain a minimally acceptable level of throughput under a wide variety of highly stressed conditions. Often these networks must operate in a multiaccess environment and must operate in the presence of rapidly varying propagation conditions. There are many approaches to such survivable networks, including spread spectrum, clustering,
14 - multicast flooding, contention resolution, and various combinations of these. Research on survivable networks must integrate the communication and network aspects of the problem, as opposed to the conventional network approach of separating these aspects. The ONR is currently funding excellent work in this area, but the size of the program is very small considering its importance. The NRL is also doing excellent work here, but on a limited scale. A significant expansion of basic research on the underlying problems of survivable networks should be of highest priority. At present there is not even a conceptual basis for contrasting the various approaches to these networks. Thus, along with the need for expanded exploration of particular approaches, there is a great need for basic research to provide a cohesive framework for the integration of comunication with network control. Given the number of supposedly survivable networks currently being built, and given the primitive state of understanding of this area, the need for more basic research is particularly pressing. NEURAL SYSTEMS Significant advances are being made by neural scientists in the understanding of biological neural networks, and basic mathematical models are beginning to evolve. Scientists and engineers have even begun to implement VLSI circuits based on these models to study the behavior of these circuits in relation to the behavior of the human brain. At present, these models are significantly lacking in their ability to model the human brain, but a breakthrough in this area would constitute a major scientific discovery comparable to or even greater in magnitude than the discoveries that led to the development of the digital computer. The development of neural network models combined with the ability to implement massively parallel systems in VLSI technology could lead to new forms of computation in which pattern recognition and feature extraction with imperfect information become more nearly feasible computationally and more reliable. With the rapid advances currently taking place in VLSI circuit technology and the neural sciences, the Navy could reap substantial benefits through the funding of interdisciplinary research projects between the electronics and biological sciences communities. Thus, we are pleased that the Electronics Division of the CRD has proposed an Accelerated Research Initiative (ARI) research program in this area in FY 1989 to: (a) promote a transfer of information between the two communities, (b) develop mathematical models and algorithms for neural networks in order to gain a deeper understanding of their behavior, and (c) implement these models in VLSI circuit technology and study their behavior. In the future, implementation of these algorithms in opto- electronic systems could be feasible. We also applaud the preliminary work at the NRL in this area. However, caution is advised. This is an extremely high-risk area in which it is difficult to make judgments about the quality of the research. Furthermore, major breakthroughs in the short term may not be forthcoming.
15 The goal of this research should be to develop new forms of algo- rithms and computation for information and signal processing. A bird and a plane both fly, but they are not identical structures. Simi- larly, a VLSI.circuit may never be identical to the human brain, but this research could lead to a better understanding of the functional processes of the brain, and thus to a realization of some of these functions in electronic systems.
- L6 -
