neuron-request@HPLMS2.HPL.HP.COM ("Neuron-Digest Moderator Peter Marvit") (09/28/90)
Neuron Digest Thursday, 27 Sep 1990 Volume 6 : Issue 56 Today's Topics: Abstracts for MIND conference in October Send submissions, questions, address maintenance and requests for old issues to "neuron-request@hplabs.hp.com" or "{any backbone,uunet}!hplabs!neuron-request" Use "ftp" to get old issues from hplpm.hpl.hp.com (15.255.176.205). ------------------------------------------------------------ Subject: Abstracts for MIND conference in October From: lindahl@evax.arl.utexas.edu (Charlie Lindahl) Date: Fri, 21 Sep 90 16:17:12 -0500 Below are the abstracts for the presentations for the MIND conference October 4-7 in Dallas. Charlie S. Lindahl Automation and Robotics Research Institute University of Texas at Arlington Internet EMAIL: lindahl@evax.arl.utexas.edu Oral Presentations DATA REPRESENTATION, NEURAL NETWORKS, AND HYBRID COMPUTATION James Anderson Dept. of Cognitive and Linguistic Science Brown University The human brain seems to be a hybrid computer, partly a highly efficient, highly evolved, massively parallel "intuitive" and "perceptual" processor, and partly a slow and inaccurate serial symbol manipulator. Neural network models seem compatible with the massively parallel hardware of the nervous system, and are probably the best current model for the parallel part of the hybrid. Traditional symbol based artificial intelligence has usually based its models on the presumed properties of the symbol manipulator. Potentially, the strengths and weaknesses of the two parts of the hyubrid compelment each other nicely, and ultimately a computer could make good use of both strategies simultaneously. However, the coupling between them in humans is still uneasy, perhaps because this novel combination is biologically recent. Let us discuss two examples of distributed hybrid computation and its associated problems of data representation as examples. First, we will discuss a network for determining large scale object motion from small scale, ambiguous local motion signals suggested by Margaret Sereno in her Ph. D. thesis. single units early in the visual system have small receptive fields, somewhat like looking at the visual system through a bunch of straws. The visual system puts together information from many sensors to give rise to perception of the large scale structure in the visual world. Determination of object motion from local motion singals is one well studied example of such a computation. The task of detecting object motion from local motion is known as the aperture problem. A very simple neural net can combine local information about edge movements to determine larger scale object motion. Second, we have recently looked at learning simple arithmetic with a neural network. We might expect neural networks to perform arithmetic poorly, because humans find it hard, and this seems to be the case. Arithmetic learning has been studied for many years. It can be viewed as only a set of arbitrary relations between symbols, for example, memorization of flashcard information. But there are richer meanings attached to the numbers that are being added or multiplied. Experiments show that humans seem to possess an internal representation of numbers related to their magnitudes. Neural network simulations obtained best results when formally redundant hybrid representations with both magnitude information and arbitrary number codes were used. Data on human error patterns in arithmetic learning indicates most errors are similar in magnitude to the correct answers. The simulations show that the model estimates values that it has not learned, signs of a rudimentary "intuition" process and even the errors observed in the simulations showed a similar pattern. INFORMATION REPRESENTATION BY PATTERNS OF EVENT-RELATED POTENTIALS IN HUMANS Jean-Paul Banquet Hospital de la Salpetriere, Paris and Center for Adaptive Systems Boston University (To be added) AN ANALOGY-BASED REASONING APPROACH TO THE CONNECTIONS VERSUS SYMBOLS DEBATE John Barnden Dept. of Computer Science New Mexico State University Symbol manipulation as used in traditional artificial intelligence has been criticized by neural net researchers for being excessively inflexible and sequential. On the other hand, the application of connectionist techniques to the types of high-level cognitive processing studied in traditional artificial intelligence presents major problems as well. I claim that a promising way out of this impasse is to build connectionist models that implement massively parallel analogy-based reasoning. This is because analogy-based reasoning avoids many of the rigidity problems leveled at traditional artificial intelligence. Although the type of system proposed is partially a version of implementational connectionism, it can also make good use of connectionist techniques for approximate matching and associative retrieval. I am accordingly modifying an existing connectionist system (Conposit), which implements standard rule-based reasoning, into a massively parallel case-based reasoning version. However, the talk will concentrate on motivations and justifications, and will not go into the new system itself in any detail. (This work has been supported in part by grant AFOSR-88-0215 from the Air Force Office of Scientific Research and grant NAGW-1592 under the Innovative Research Program of the NASA Office of Space Science and Applications.) USING GENETIC ALGORITHMS IN THE DESIGN AND CONSTRUCTION OF ARTIFICIAL NEURAL NETS Warren Bean Texas Instruments, Inc. Genetic algorithms provide a means for discovering efficient topologies, training set partitions, and parameter settings for use in the design and construction of neural nets. In this talk I will illustrate some of the ways genetic algorithms can be used to design or optimize the process of finding good neural network solutions to certain problems. I will review some of the hybrid approaches in which a genetic algorithm is combined with a neural network to solve problems that might have been difficult to solve by just one technique alone. I will compare a Holland classifier algorithm in which production rules are discovered via a genetic algorithm with a specific implementation of a neural net. By this comparison I will try to highlight the features each approach brings to bear in solving problems in adaptation. THE BRAIN AS A COMPUTER -- OR IS IT? Claude Cruz Plexus Systems, Inc. The operation of nervous systems has been likened to many things in the past. To Galen, the nervous system was a pneumatic network. Descartes thought of the brain as a fantastic clockwork. More recently, the human brain has been compared to a telephone exchange. In this paper, we will view the operation of an advanced nervous system in light of the "paradigm du jour": the computer. This comparison will be undertaken from a functional perspective. That is, we will focus on the information-processing properties and capabilities of both computers and nervous systems. Little will be said of the possible ties to neurobiology, apart from a commonality of some high-level operating principles. Likewise, we will not discuss the circuit-level details of computer implementation approaches. In keeping with the title of this Workshop, we will delineate the fundamental operations performed by a computer. We will explore what the notion of representation means in the context of a computer, and in the context of a nervous system. We will see that both computers and nervous systems may be viewed as pattern- processing mechanisms, though of rather different kinds. Possible analogous mechanisms in nervous systems will be identified, and we will also note differences in how computers operate. These considerations will lead to a proposal that we distinguish between "symbolic" processing systems (such as computers) and "semantic" processing systems (nervous systems). QUANTUM NEURODYNAMICS AND THE REPRESENTATION OF KNOWLEDGE Robert Dawes Martingale Research Corporation Quantum Neurodynamics (QND) is the theoretical basis for a neurocomputing architecture which implements powerful stochastic filtering methods. Unlike the vast majority of existing artificial neural system architectures, the QND architecture is inherently designed to comprehend nonstationary observations of nonlinear dynamical systems. Simulations have shown that a QND observer can simultaneously track multiple accelerating targets in simulated imagery and can solve the "broom balancing" problems from well into its nonlinear region. This talk explores the implications of the QND model for knowledge representation in biological systems. We show how QND can model representational dysfunctions resulting in multiple personality, and we introduce models for the reification of synthetic inference and deductive logic. AN EXPERT SYSTEM FOR THE DIAGNOSIS OF PAPULOSQUAMOUS SKIN DISORDERS USING ID3 AND NEURAL NETS Jonathan Fenner, Kyle Morris, and Lynn Peterson Dept. of Computer Science University of Texas at Arlington YoungOhc Yoon Dept. of Computer and Information Science Southwestern Missouri State University This paper presents the results of the development of an expert system to assist medical students in diagnosing papulosquamous skin disorders. These disorders form a broad category of skin disease which is fairly common, frequently confused, and have widely varying etiologies. This expert system includes two methodologies running in parallel: a connectionist approach, and a symbolic learning system based on ID3. This paper will compare both approaches in terms of accuracy, learning time, and sensitivity to learning set size. TOWARD A MARKOV RANDOM FIELD MODEL OF STORY COMPREHENSION AND RECALL Richard Golden School of Human Development University of Texas at Dallas David Rumelhart Department of Psychology Stanford University Understanding and recalling a story is formally viewed as an optimal control problem. In particular, the reader's mental state is modelled as a collection of d binary-valued features. This state is represented as a point in a d+1-dimensional situation state space by the state vector X(t). Thus, a sequence of points (or trajectory) in situation state space is assumed to reference some space-time process in the physical world. The reader's knowledge of the world consists of trajectories (i.e.,sequences of situations) in situation state space. The knowledge base is compactly represented by a subjective probability distribution of the form p(X(1), ..., X(T)~W) which expresses the reader's belief that trajectory X(1), ..., X(T) will occur given that the reader is in memory state W. The reader also possesses some unmodifiable "long-term" beliefs about W which are embodied in the prior distribution p(W). A story is modelled as a set, D, of constraints in situation state space. In particular, let D be a particular description of a story. The reader's belief that D is a description of a trajectory X(1), ..., X(T) is modelled by the subjective probability distribution P(D~X(1), ..., X(T)). The comprehension and recall processes are assumed to operate as follows. The reader computes the "most believable" situation state space trajectory given a story description D* by maximizing the likelihood function p(D*~X(1), ..., X(T))p(X(1), ..., X(T)~W)p(W) with respect to X(1), ..., X(T) and W. In the process of this computation the belief that feature xi of situation X(t) is active, p(xi(t) = 1~W, ~xj~xi), is also computed. Let ~c be the system comprehension threshold and ~r be the system recall threshold where 0<~c<~r. If p(xi=1~W, ~xj(t)~xi(t))>~c, then xi(t) is "understood" by the system to be active in situation X(t). If p(xi=1~W, ~xj(t)~xi(t))>~r, then xi(t) is defined as "recalled" by the system. Some preliminary simulations which illustrate the viability of this approach are described. In the simulations, a story grammar analysis of several short children's stories is used to derive the situation state space representation. Next, a Markov random field modelling approach is used to construct consistent computationally tractable probability distributions through the used of meaningful independence assumptions. Finally, the resulting probability distributions are optimized using a gradient descent algorithm which is formally equivalent to a special type of recurrent neural network algorithm. KNOWLEDGE REPRESENTATION IN THE HUMAN VISUAL SYSTEM William Hudspeth Center for Brain Research and Info. Science Radford University Visual percepts and concepts are organized in a way that various objects are classified according to similarities of attributes and meaning. This property can be defined by a multidimensional metric, and visual stimuli can be designed to assure classification by these principles. In the present work, perceptual and conceptual classification stimulus sets were used to obtain visual evoked potential recordings. The quantitative relationships among the stimuli in each set were computed as an a priori model of how humans would classify perceptual and conceptual stimulus sets. This model predicted the relationship among visual evoked potentials within each stimulus set with 99% accuracy. The results of these studies show that the brain electrical recordings, and their numerical indices, retained all of the information specified in the perceptual and conceptual classification specified by the a priori model. Thus, the human visual system provides a linear transfer function for the relationships among stimulus objects according to their attributes and meanings, showing that the electrical activity of the human brain specifies the exact attributes of objects in the visual world. KNOWLEDGE AND THE STRUCTURE OF MIND Samuel Leven Center for Brain Research and Info. Science Radford University (To be added) CONTINUOUS SYMBOL SYSTEMS: THE LOGIC OF CONNECTIONISM Bruce MacLennan Dept. of Computer Science University of Tennessee, Knoxville It has been long assumed that knowledge and thought are most naturally represented as discrete symbol systems (calculi). Thus a major contribution of connectionism is that it provides an alternative model of knowledge and cognition that avoids many of the limitations of the traditional approach. But what idea serves for connectionism the same unifying role that the idea of a calculus served for the traditional theories? It is the idea of a continuous symbol system. This talk presents a preliminary formulation of continuous symbol systems and indicates how they may aid the understanding and development of connectionist theories. It begins with a brief phenomenological analysis of the discrete and continuous; the aim of this analysis is to directly contrast the two kinds of symbols systems and show their equal primacy. Next, based on the phenomenological analysis and on other observations of existing continuous symbol systems and connectionist models, I sketch a mathematical characterizationn of these systems. Finally the talk turns to some applications of the theory and to its implications for decidability issues and the theory of computation. NOVELTY MONITORING, METACOGNITION AND CONTROL IN A COMPOSITE HOLOGRAPHIC ASSOCIATIVE RECALL MODEL (CHARM): IMPLICATIONS FOR KORSAKOFF AMNESIA Janet Metcalfe Department of Psychology Dartmouth College This work derives from a technical problem in composite-trace distributed models of human memory, and particularly in CHARM. The problem is of interest because it appears that with certain kinds of organic brain disorders, people experience a similar problem. Briefly, the composite trace employed as a central construct in such models can become catastrophically out of control. To solve the problem a pre-storage 'novelty-familiarity' monitor and a simple control procedure need to be implemented. Experimental evidence converging on the idea that output from such a monitor underlies people's metacognitive judgments of 'feeling of knowing' is outlined. Breakdown of the monitoring-control mechanism is shown to produce Korsakoff-like symptoms in the model. Impairments in feeling-of-knowing judgments and the failure to release from proactive inhibition, both characteristic of Korsakoffs, are, thus, attributed to monitoring-control failure rather than to deficits in the basic memory system. THE ADAPTRODE: AN ADAPTIVE SIGNAL PROCESSOR THAT CAN ENCODE ASSOCIATIONS PROVIDES A NEW LEARNING MECHANISM FOR NEURAL NETWORKS George Mobus Department of Computer Science University of North Texas Learning rules in artificial neural networks have been formulated with an underlying assumption that associativity is the primal feature of learning phenomena. Thus all such rules attempt to form associations either between an input signal and the output of the neural element (Hebb Rule), between two adjacent input signals (Alkon's Local Interaction) or between an output and a desired output (delta rule and derivatives). Without debating the significance of associativity in learning, I assert that information encoding based strictly on associativity loses information. The assumption of associativity leas to highly constrained rules for changing synaptic efficacy and tends to exclude time. The Adaptrode is an adaptive signal processor that encodes information in response to its experience in multiple time sdcales. Its behavior mimics both synaptic junctions and the axonal hillock in a variety of neuronal types. A simple "permission" gating scheme provides a means for establishing temporal and spatial associations among Adaptrodes. Thus in neuromimes using Adaptrode elements as synapses and thresholds, it it possible to model Hebbian, Local Interaction and error minimization learning. But in addition, associations can be established between different Adaptrodes at different time scales. Short and long-term potentiation and neuromodulated memory phenomena have been modeled. A simple robot decision controller demonstrates conditioned associations between conditionable and unconditionable stimuli. REPRESENTING PROPOSITIONAL LOGIC IN INTERACTIVE SYMMETRIC NETWORKS Gadi Pinkas Dept. of Computer Science Washington University, St. Louis Interactive symmetric connectionist models, like the Hopfield model and Boltzmann machine, use gradient descent to minimize energy functions. We show an efficient way to translate a given propositional Well Formed Formula (WFF) into an energy function that represents the WFF. The energy function can then be implemented using a connectionist network that minimizes it. The technique we use generates an energy function such that the set of vectors that minimizes the function (to global minima) is equal to the set of truth assignments (models) that satisfy the WFF. The connectionist network can therefore be seen as performing a search for a satisfying model. We found an equivalence between the problem of satisfiability and the problem of connectionist energy minimization. For every WFF we can find a class of energy functions that represent the WFF and for every connectionist energy function we can find a class of WFFs that are represented by the function. The algorithm to transform a WFF into a connectionist network produces a network that may include hidden units and determines its exact topology and weights. The network that is generated is efficient in the sense that the number of hidden units created is linear in the length of the original WFF and the fan-out of these hidden units is bounded by a constant. We also show that by using sigma-pi units we cna eliminate completely the need for any hidden units. Given a set of possibly contradicting WFFs, the technique can be used to find the maximal consistent subset of a set of beliefs. Each WFF may have a penalty assigned to it that is effective if the WFF is not satisfied. The gnerated network searches for a model that minimizes the total penalty. In the special case where all the WFFs have the same penalty, the system minimizes the cardinality of the subset of unsatisfied WFFs. At equilbrium those WFFs that are satisfied by the minimum penalty model are in the maximal consistent subset. Finding the maximal consistent subset may be applied to certain non-monotonic paradigms (Truth Maintenance Systems for example) where we can consider a total order of preference among the beliefs. In such a system for example, absolute logic rules should be preferred over defeasible rules and assigned a higher penalty. Contradicting arguments compete among themselves and those who win are the ones that cause the less damage in the sense of total penalty or cardinality. Such connectionist non-monotonic systems feature incremental updating of the knowledge so that when we add to or update the knowledge base, we change the network locally without the need for re-calculating it all over again. The same method can also be used for tolerating possible observation errors by ignoring those which contradict a consistent majority set. Adding a little more sophistication, we can now implement a 3-value logic propositional inference engine. We want the network to deduce "unknown" when a proposition can be either "true" or "false" in two different satisfying models of the maximal consistent subset. For the rest of the poropstions the network is forced to deduce either "true" or "false." The trick is to construct an energy function with only one global minimum by introducing weak constraints that cause the network to prefer "deducing unknown" as long as it does not violate any fo the 3-value logic WFFs that are in the maximal consistent subset. RECURSIVE DISTRIBUTED REPRESENTATIONS Jordan Pollack Dept. of Computer and Information Science Ohio State University A longstanding difficulty for connectionist modeling has been how to represent variable-sized recursive data structures, such as trees and lists, in fixed-width patterns. I present a connectionist architecture which automatically develops compact distributed representations for such compositional structures, as well as efficient accessing mechanisms for them. Fixed-dimensional patterns which stand for the internal nodes of finite sets of fixed-valence trees are devised through the recursive use of back-propagation on three-layer auto-associative encoder networks. The resulting representations are novel, in that they combine apparently immiscible aspects of features, pointers, and symbol structures. They form a bridge between the data structures necessary for high-level cognitive tasks and the associative, pattern recognition machinery provided by neural networks. THE NEURODYNAMICS OF FRONTAL LOBE PROCESSING Karl Pribram Center for Brain Research and Info. Science Radford University Evidence will be presented that (a) the effects of amygdalectomy on serial position, (b) the effects of hippocampectomy on primacy, and (c) the effects of far frontal lobe resections on intralist interference (relative recency) stem from inadequate perceptual processing at the time of initial exposure to the list of items, the establishment of an episode, the context that stabilizes further processing, and not to effects on the trace of the sensory input. We have all experienced a related phenomenon when we attempt to recite a poem or rehearse a melody: should we be interrupted or fail, for the moment, to be able to continue the recitation or rehearsal we often find it necessary to begin again at the beginning of the entire poem or piece, or at least at the beginning of a major section. Murdock has developed a convolutional model that better handles these serial position effects than does a matrix model (which deals more readily with categorical processing of prototypes). Extensions of the convolutional model by Smolensky (a dynamical theory) and Pribram (a holonomic brain theory) account for other behavioral and experiential phenomena linked to the functions of the frontolimbic forebrain. Processing that leads to practical inference is one such function. A mathematical model of the neurodynamics that can lead to inference will be described. EXPLORING CONNECTIONIST APPROACHES TO LEGAL DECISION MAKING Wullianalur Raghupathi and Lawrence Schkade Department of Information Systems University of Texas at Arlington Raju Bapi and Daniel Levine Department of Mathematics University of Texas at Arlington In this exploratory paper, we discuss research in connectionism and neurophysiology and examine the potential implications for understanding and modeling legal decision making. Artificial Intelligence approaches such as rule-based and case-based, while adequate for implementing systems in specific legal domains, have proved insufficient for modeling complex, large scale legal systems that assimilate a variety of inputs. This paper contributes to three important goals: modeling legal decision processes as special instances of human decision making, conceptualization of intelligent, computer-based legal systems, and moving toward integrated frameworks for research. - ------------------------------------------------------------- FROM SIMPLE ASSOCIATIONS TO SYSTEMATIC REASONING Lokendra Shastri Dept. of Computer and Information Science University of Pennsylvania Human agents are capable of representing a large body of structured and systematic knowledge and drawing a variety of inferences based on this knowledge with extreme efficiency. These inferences are by no means trivial and support a broad range of cognitive activity such as understanding language and performing commonsense reasoning. We present work on connectionist knowledge epresentation and reasoning systems that is a step toward a computational account of this remarkable reasoning ability. A key technical problem that must be solved in order to represent and reason with structured knowledge is the dynamic creation and propagation of bindings. We show how this problem may be solved using simple phase-sensitive oscillatory neuronal elements that allow the system to maintain and propagate a large number of dynamic bindings simultaneously. Poster presentations ARCHITECTURES FOR UNCERTAIN REASONING James Austin, Thomas Jackson, and Alan Wood Dept. of Computer Science University of York, England This paper presents the basis of a neurocomputer architecture for uncertain reasoning. The arguments for applying neural networks to this domain are discussed together with the broad architecture of the system. The work considers the boarder range aspects of a neural computer architecture, not just the implementation of the neural network. It is shown how a two stage system comprising of a front end SIMD processor and a tightly coupled back end neural processor is to be used in initial investigations. A neural associative architecture is considered for the back end processor that is based upon the ADAM system, a high speed extensible network developed at York. The major issues that relate to the full development of the system are presented, along with an indication of the applications of such a system. - ------------------------------ SETS, TUPLES, GRAPHS, PROPOSTIONS, AND FINITE REGULAR LANGUAGES IN A HOPFIELD-STYLE NETWORK Arun Jagota Dept. of Computer Science State University of New York at Buffalo A Hopfield-style one-layer network has been developed [1] whose stable states are the maximal cliques in the underlying graph, allowing analysis and proof of many of its properties using graph-theoretic techniques. This network with graph-theoretic properties is also found to be superior to any known one-layer network model as a content-addressable memory/information storage device in it's space/time complexity and storage capacity. In particular, it has exponential storage capacity, although spurious memories do develop. As importantly, it's relationship with graph theory, in particular the relationship between its stable states and maximal cliques has established it as a maximal clique machine. In other words, any existing or not-as-yet-discovered reductions of problems to maximal cliques in graphs also implies that such problems can be mapped onto the network. At this workshop, we will describe five key problems that can be mapped onto the network, storage and retrieval of sets (for CAM), storage and retrieval of Tuples (for CAM and databases), storing arbitrary graphs, storing CNF Propositional formula/testing for their truth-assignments and storing finite-regular languages/testing for membership. The ability of the network on these problems and their refinements and it's near-optimal space and time efficiency may have important consequences for Knowledge Representation. Moreover, the dynamic properties of the network, as proven in [1], are quite different from those of other networks and may have important consequences for knowledge retrieval from noisy or incomplete information or information containing multiple hypotheses. Storing Sets and Tuples is a direct consequence of the network Learning rule [1]. The network has significantly better CAM properties for Tuples than Sets [1]. Sets can be stored as Tuples by a suitable transformation. Any graph can be stored in the network by presenting its edges as binary sets. CNF Propositional formula and finite-regular languages are stored by transforming their expressions to particular kinds of graphs and storing these graphs in the network. These transformations will be described at the workshop and it will be shown how such storage is accomplished. For a restricted class of finite regular languages, the network stores any language in this class perfectly and can test any string for membership in one time step. There are languages in this class that have an exponential number of strings. Even for such languages, the network needs only polynomial storage (number of connections) and can be trained in polynomial time. Finally, changing the value of a global network parameter allows the stable states to be subgraphs of a minimum density, where the two extremes of this global parameter value give rise to stable states that are connected components at one extreme and maximal cliques at the other[1]. This and other graph-theoretic properties are proven in [1] and their implications for clustering in general and Concept and Knowledge clustering in particular will be discussed at the workshop. And lastly, this network model has been implemented in software and applied to moderate-size real-world problems. We will also use this software on detailed examples of the problems mentioned above, specifically to obtain results which seem theoretically difficult and report such results at the workshop. We will also offer this software to any interested researchers. [1] A new Hopfield-style network for content-addressable memories, Arun Jagota (1990), Technical report TR 90-02, Department Of Computer Science, State University Of New York At Buffalo. USA SELECTIVE ATTENTION IN CONTEXTUAL MEMORY Nilendu Jani Dept. of Computer Science University of Texas at Arlington The primary purpose of this paper is to provide a connectionist model for the cognitive process of sleectively choosing a concept out of a pool of candidate concepts present in the current context. The distinction here is the need for contextual information. Initially key issues concerning the modelling of memory are identified by asking a few fundamental quesitons about blocking, interruption, suspension, resumption, attentional span, spreading activations, consciousness, recognition, retrieval, learning, retaining, and forgetting. The significance of contextual memory and selective attention in comprehensive modelling of human brain and day-to-day cognitive processing is discussed. The recent trends in neural network based architectures in modelling selective attention are identified. The one particularly of interest in this paper is the model based on competition by Grossberg and Levine, which uses drive representations for selective attention. The Grossberg-Levine model is extended to take context into account. The enhanced model concentrates more on the role of selective attention in retrieval than in learning or conditioning. Furthermore, the significance of separate attentions at the sub- system level and at the global level is identified. The existence of contextual memory, and the distinction of attentions at the sub- system and global levels, are validated and explained on intuitive grounds, not on a physiological or experimental psychological basis. A scenario depicting a typical day-to-day event is given to illustrate the nature of selective attention within contextual memory. Finally a mathematical connectionist model of selective attention is proposed, which incorporates context and different levels of attention. NEURAL NETWORK FOR HIGH-SPEED IMPLEMENTATION OF RULE-BASED SYSTEMS Arun Maskara Dept. of Computer and Information Science New Jersey Institute of Technology Andrew Noetzel Brooklyn Polytechnic University We consider the use of a neural network for the parallel selection of rules in a rule-based expert system. In our approach, the expert system rules are presented as patterns that are learned by a neural network. Each rule is learned after only one presentation. The learning mechanism is similar to that of Grossberg's ART-1 model, but the architecture has been adapted for the specific requirements of the expert system application. After learning, the neural network acts as a content-addressable memory. When presented with a partial rule description, it produces a parallel specification of all rules that may be applicable. The neural network is a multilayered, data-driven network. After choosing the applicable rules for a particular problem state, the architecture allows for the execution of multiple control paths. Inferencing is implemented by allowing feedback between the various levels of the network. A combinational logic net is designed to act as the rule interpreter. The architecture of the neural network and interpreter is demonstrated, along with examples. The results of simulation examples are shown. Possible variations and generalizations of the design are discussed. USING AN ARTIFICIAL NEURAL SYSTEM TO DETERMINE THE KNOWLEDGE BASE OF AN EXPERT SYSTEM George Whitson, Cathy Wu, and Pam Taylor Computer Science Department University of Texas at Tyler This paper gives a mapping of rule based expert systems into artificial neural expert systems. While the mapping is not one-to- one, it does show that the two systems are essentially equivalent. There are, of course, many examples of artificial neural systems that are not expert systems. We can use the reverse mapping of an artificial neural expert system to a rule based expert system to determine the knowledge base of the rule based expert system, i.e., to determine the exact nature of the rules. This yields an automated procedure for determining the knowledge base of an expert system that shows much promise. We have implemented this expert system tool on several larger microcomputers, including an Intel Sugarcube. The Sugarcube implementation is a very natural one. ------------------------------ End of Neuron Digest [Volume 6 Issue 56] ****************************************