Proceedings of the Annual Meeting of the Cognitive Science Society

Concepts can be represented by distributed patterns of activity in networks of neuron-like units. One advantage of this kind of representation is that it leads to automatic generalization. When the weights in the network are changed to incorporate new knowledge about one concept, the changes affect the knowledge associated with other concepts that are represented by similar activity patterns. There have been numerous demonstrations of sensible generalization which have depended on the experimenter choosing appropriately similar patterns for different concepts. This paper shows how the network can be made to choose the patterns itself when shown a set of propositions that use the concepts. It chooses patterns which make explicit the underlying features that are only implicit in the propositions it is shown.

The present paper is a preliminary report of our work exploring skill learning and repetition priming in parallel studies of mirror symmetry detection in humans and network models. The memory mechanisms supporting the acquisition of skill and repetition priming in humans have been the subject of much speculation. On one account, drawing on the distinction between procedural and declarative learning, these learning phenomena grow out of experience-based tuning and reorganization of processing modules engaged by performance in a given domain, in a manner that is intimately tied to the operation of those modules. Such learning appears similar to that suggested by the Incremental learning algorithms currently being explored in massively-parallel connectionist models (e.g., the Boltzmann machine). In the present work, both learning phenomena were observed in the behavioral data from human subjects and the simulation data from the network models. The network models showed priming effects from the start of de novo learning despite being designed to handle generalization to new materials - the essence of skill learning - and without additional mechanisms designed to provide a temporary advantage for recently presented material. Priming occurred for the human subjects despite the use of novel materials for which pre-existing representations cannot already be present in memory. These findings support the notion that skill learning and repetition priming are linked to basic incremental learning mechanisms that serve to configure and reorganize processing modules engaged by experience.

Qualitative judgment, the ability to evaluate attributes that imply some degree of 'goodness' or preference, poses important problems for the information processing paradigm. In this paper one form of qualitative judgment, contextual judgment of magnitude, is analysed in some detail. The results of psychophysical experiments are consistent with the idea that human magnitude representation is based on a contextual coding process in which an actual stimulus is compared with a sample of traces of previously encountered similar stimuli. Such a coding process is hard to realize in a conventional memory system. A distributed model for contextual magnitude judgment is described, in which this trace sampling process is feasible, when special provisions for the use of resonance information are made. Resonance coding involves the representation within a memory system of the memory activity caused by specific patterns of stimulation. A possible implementation of resonance coding, detection of dissonance, is briefly described. The hypothesis is put forward that evaluation of memory resonance plays an equally important role in other forms of qualitative judgment.

NETtalk is a massively-parallel network that leams to convert English text to phonemes. In NETtalk, the memory representations are shared among many processing units, and these representations are learned by practice. In humans, distributed practice is more effective for longterm retention than massed practice, and we wondered whether learning in NETtalk had similar properties. NETtalk was tested on cued paired-associate recall using nonwords as stimuli. Retention of these target items was measured as a function of spacing, or the number of interspersed items between successive repetitions of the target. A significant advantage for spaced or distributed items was found for spacings of up to forty intervening items when tested at a retention interval of 64 items. Conversely, a significant advantage for massed items was found if testing immediately followed study. These results are strikingly similar to the results of many experiments using human subjects and suggest an explanation based on distributed representations in massively-parallel network architectures.

A primary constraint on using a pronominal anaphor is that it must agree with its antecedent in number. However, there are situations in which pronouns act as conceptual anaphors. For example, in the discourse, "I think I'll order a frozen margarita. I just love them.", the pronoun "them" does not refer to a single margarita, but perhaps all the margarltas the speaker has ever tasted. When anaphors operate in this way, they are often mismatched with their ILteral antecedent in number. Three situtations when conceptual anaphora occurs are identified: when referring to the members of a Collective Set (as opposed to the set per se), a Multiply occurring Item or Event (v/ersus a Unique Item/Event), or a Generic Type (versus a Specific Token). Two experiments are reported. The first demonstrated that subjects consider a mismatched, plural pronoun more natural than a matched, singular pronoun when it follows a Collective Set, Multiple Item/Event, or Generic Type noun. Conversely, subjects consider a matched, singular pronoun more natural when it follows an Individual Member of a set, Unique Item/Event, or Specific Token noun. The second experiment demonstrated that subjects comprehend a mismatched, plural pronoun faster than a matched, singular pronoun when it follows a Collective Set, Multiple Item/Event, or Generic Type noun, but they comprehend a matched, singular pronoun faster when it follows an Individual Member, Unique Item/Event, or Specific Token noun. This suggests that when comprehenders encounter conceptual—though mismatched anaphors—they do not have to reinstate the multiple entities into their mental representations.

Humans and animals have the ability to learn complicated configurations of environmental cues that are predictive of important events. In cljissical conditioning, this task is called configural conditioning. Psychologists have studied this phenomenon since Pavlov's time, yet several of the contemporary learning models provide only partially satisfactory explanations. Most models provide mechanisms which select among possible predictive stimuli, but they fail to explicitly identify predictive combinations of stimuli and are thus restricted to learning only a relatively simple set of possible associations. In this paper we discuss a learning method which accounts for some configural conditioning results. Using an implemented system, we demonstrate the effectiveness of this method by modeling configural conditioning data from a pair of representative experimental studies.

Networks composed of layers of adaptive elements provide a rigorous explanation for complex associative learning phenomena. In particular, a network composed of three adaptive elements can explain previously intractable phenomena, namely the rapid rate of reacquisitions, learning-to-learn, spontaneous configuration, and negative patterning (the exclusive-OR problem). This paper will compare the results of computer simulations to the behavioral results of classical conditioning experiments using the rabbit's nictitating membrane response.

Sutton-Barto (SB) model of learning is based on a neuron-like adaptive element. The model has computational features suitable for describins a variety of classical conditioning phenomena, including blocking, conditioned inhibition, and higher-order conditioning. However, it presently does not describe within-trial phenomena related to conditioned response (CR) topography. W e here describe in detail an extension of the SB element, referred to as the Sutton-Barto-Desmond (SBD) model, which is capable of simulating topography of the conditioned nictitating membrane response (NMR) of the rabbit. The SBD model places certain constraints on the SB modePs parameters and makes some additonal assumptions about the form of inputs to the element. The model describes (1) the gradually increasing amplitude of the C R within a trial with the peak amplitude at the temporal locus of the US, (2) the decrease in C R onset latency over training, and (3) appropriate interstimulus interval (ISI) functions, with optimal learning occurring with an ISI of .25 seconds. In addition, the model lends itself to descriptions of neuronal firing related to the CR. W e believe the S B D model may have implications for neurobiological studies of learning and memory.

cognitive agent uses representations to reason about the world. An "introspective" cognitive agent has the ability to manipulate representations (meta representations) of its own representations. If such an agent were to embody its beliefs in its representations, then the agent could reason about its own beliefs by manipulating its meta representations. A "belief reasoner" can reason about the beliefs of other agents. There has been considerable research in the construction of belief reasoners. This paper observes that the construction of such systems can, in large part, be reduced to the task of constructing introspective systems. We illustrate how an introspective agent can use analogy-based reasoning to construct an architecture for belief reasoning on the basis of examining its own architecture.

Causality plays an important role in human thinking. Yet we are far from having a complete account of causal reasoning. This paper presents an analysis of causal reasoning about changes in quantities. W e abstract from AI theories of qualitative physics three dimensions along which causal reasoning about quantities may be decomposed. W e then use this framework to make some psychological predictions.

presented indicating that spontaneously generated analogies can play a significant role in expert problem solving. Since not all analogies are valid, it is important for the subject to have a way to evaluate their validity. Three methods for evaluating analogical validity are identified using observations from thinking aloud problem solving protocols as well as examples from Newton and Galileo. In particular, this paper focuses on an evaluation strategy called bridging that has been observed in solutions to both science and mathematics problems. In constructing a bridge, the subject finds an intermediate case that is seen as "in between" the analogous case and the problem situation because it shares important features of both. Many of the bridges observed appeared to be novel inventions created by the subject. These empirical studies have led to the construction of a more detailed theory for how analogies can be used effectively in instruction. Some of the strategies observed in experts appear to have high potential for helping science students overcome persistent misconceptions in the classroom.

In psychophysics, the effects of light backgrounds and photopigment bleaching can be equated in a variety of situations involving sensitivity, flicker, and the subjective perception of brightness. We have investigated such possible equivalences at the single unit level through intracellular recordings from vertebrate rods. At a given level of background or bleaching adaptation, vertebrate rods respond with characteristic waveforms to brief flashes of light of increasing intensity. The peak responses can be plotted vs. flash intensity to give a response curve in each state of adaptation. Two consequences of light adaptation are a shift to higher light intensities and a compression of the response curve. We have been able to establish equivalences in single units between bleaching and backgrounds in terms of threshol delevation and response compression. Moreover, we have found that such equivalent (background, bleaching) pairs have similar response curves. But these response curves are direct measures of stimulus flash intensity . Based on the close parallels between the psychophysical and the neurophysio 1 ogica 1 equivalences of bleaching and backgrounds, we therefore propose that brightnessperception is mediated by the peak responses of photoreceptors. It is not often that behavioral observations can be traced back to the neuronal machinery and the latter suggest psychophysical tests which can further clarify brightness equivalence relations. This approach is discussed in the light of our results.

The role of midbrain visual centers for orienting attention was studied in chronometric experiments measuring the effect of pre-cues on simple reaction time to detect a peripheral luminance change. Two types of cues were tested: Exogenous cues—a peripheral luminance change which did not predict target location; and Endogenous cues—a central arrow vrfiich predicted the likely target location. Patients with peri-tectal midbrain degeneration from progressive supranuclear palsy showed deficits in orienting to both types of cues. In normal human subjects tested monocularly, we compared orienting into the tenporal hemifield which has more direct access to the midbrain superior colliculus) with orienting into the nasal hemifield. Exogenous cues produced equivalent speeding of detection at cued locations in both hemifields; but nasal cues produced more slowing of detection at uncued locations. Endogenous nasal cues produced earlier speeding of detection at cued locations than temporal cues; and at later intervals, they produced more slowing of detection at uncued locations. Both cortical and subcortical visual systems appear to be integrated in orienting to both exogenous and endogenous information. Whereas the subcortical pathway receives input mainly from the temporal hemifield, the cortical system is biased in orienting to the nasal hemifield; and its committment produces moreslowing of detection at unattended locations.

Neuropsychological studies indicate that hippocampal formation in the human medial temporal lobe (MTL) plays a crucial role in the formation and retrieval of memories for recent events. We have found that components of neural activity recorded from the MTL during recognition memory testing discriminate between repeated and nonrepeated words. Such recordings are made possible via intracranial electrodes implanted for the isolation of seizure foci in epileptic patients. Similar activity can be elicited during lexical decision tasks, where it is also sensitive to stimulus repetition. It has been suggested that repetition effects on lexical decision performance measures are a reflection of procedural learning, unlike the type of learning that underlies recognition memory performance, does not involve the MTL. However similar changes in the MTL response during both lexical decision and recognition memory suggests that a common mechanism contributes to repetition effects across tasks. These potentials provide a juncture for studies at both the psychological and synaptic levels of analysis.

Criticisms of probability as being epistenxjlogically inadequate as a basis for reasoning under uncertainty in Al and rule-based expert systems are largely misplaced. Probabilistic schemes appear to be the best way to deal with dependent evidence, and to properly combine diagnostic and predictive inference. Suggestions that expert systems should duplicate human inference strategies, with their documented biases, seem ill-advised. There is evidence that popular schemes perform quite poorly under some circumstances and there is an urgent need for careful study of when they can be relied upon. Some promising probabilistic alternatives are available, but they need to be demonstrated in realistic applications.

What are the factors that cause a problem solver to become blocked? And what are the factors that allow a person to become unblocked? These are the motivating questions for a set of studies we conducted of individual and joint problem solving. By constructing an isomorph of the classic "water jar" problems (Luchins, 1942) as a dynamic graphic microworld, we were able to identify several factors involved in producing blocked states. By comparing the behavior of individuals tackling the "missionaries and cannibals" problem to pairs of people solving this problem, we have been able to identify ways in which problem solvers operating in a social context are able to overcome problem solving blocks that are difficult for individuals. These studies point to the importance of "reflection" (evaluation of problem solving results) for flexible problem solving. These results may also account for the difficulty in showing learning in "discovery learning" uses of computers, such as the use of Logo, since such uses also often do not encourage students to reflect on the outcome of their problem solving.

We describe the initial prototypes of several Intelligent tutoring systems designed to build students' scientific Inquiry skills. These Inquiry skills are taught In the context of acquiring knowledge of principles from a mlcroworld that models a specific domain. We have Implemented microworlds for microeconomics, electricity, and light refraction. All of the systems are highly interactive; students can pose questions, conduct experiments by manipulating domain specific factors, and record results. Using protocol studies of expert and non-expert learners using these microworlds we identify important inquiry strategies. W e have represented these strategies formally, allowing the mlcroworld to detect effective and ineffective Inquiry strategies. W e conclude with a description of a partially implemented 'inquiry coach*. This coach will be incorporated into the microworlds and teach the Inquiry strategies in the context of the specific mlcroworld domain knowledge.

The construction and retrieval of plan schemas was studied using a programming task. Novice programmers were asked to write programs and their problem solving strategies analysed from the manner in which they expanded a program goal. Four programs were used in the study, comprising two sets of problem isomorphs. These isomorphs required the same programming plans for solution, but differed in their cover stories. Plan development withm a prob em solving episode and across episodes was thus easily tracked. Development of a plan to achieve a problem goal showed a focus of attention on the current goal to the exclusion of concerns about other parts of the program. A solution was not found by a process of reasoning and top-down design. Rather, one goal was selected for expansion and in the process of this expansion new goals were discovered and solved to the extent required by the current context. The program was built up from solving individual goals, not down from the problem specification. The development of plans to achieve three of these goals, the selection, simple (non-looping) sum and read loop goals, is discussed in some detail to show the relation between problem solving and schema formation. Evidence of top-down design using schemas appeared only late in the study, suggesting that its use in teaching should be delayed until the novice can 'speak the language' of schema retrieval.

Our goal is the complete integration of natural language understanding with the rest of cognition. Two mechanisms that we have developed and implemented to achieve this goal are: (l) the Direct Memory Access Parsing (DMAP) algorithm, based on the notion of lexically-guided memory search and concept refinement, and (2) an inference-triggering process based on the notion of concept refinement failures. Together, these two mechanisms form a tightly-integrated system of parsing and inferencing, with no artificial boundaries between them.

The model discussed here is offered as a prototype of the use of a computational model to explore alternate hypotheses and to suggest possible answers to some of the questions which have been addressed in the study of language acquisition. Why does not the child end up with an overly generalized grammar or lexicon? There is much evidence concerning the kinds of generalizations and over-generalizations that children make. However if we permit no overt and specific correction of the child's errors, then how is it that errors of over-generalization do not persist into adult speech? One answer to this question is proffered by attaching a system of weights to hypotheses. There are two related problems to be solved. Some mechanism in the model must allow erroneous hypotheses to be corrected; in addition there must be a way that more mature constructs can replace earlier ones. The model accomplishes these two tasks by means of a system of weights which represent confidence values and recency values. By this system more frequently matched constructs are preferred over less frequently matched constructs, and more recent hypotheses are favored for testing. This learning paradigm is illustrated by a set of procedures for learning the past tense of verbs in English. The scheme has the advantage that for a period of time when confidence factors are approximately in balance two or more constructs can co-exist. Thus we need not talk of rules or individual cases which have been learned or have not yet been learned but rather of a continuum in which rule schemas are either strong or weak.

In this talk I will present the basic elements of a computational theory of discourse structure.^ A proper account of discourse structure is needed both as the basis of an account of discourse meaning (a semantic task) and to underlie a model of discourse processing. It provides the former by specifying the basic units a discourse comprises and the ways in which they can relate. It plays a key role in discourse processing by stipulating constraints on those portions of a discourse to which any given utterance in the discourse must be related.

recent years, much research has been aimed at the study of episodic reminding and its relationship to the process of understanding in both man and machines. Because of the ill defined nature of the functionality of episodic remindings in the understanding process, however, little progress has been made in uncovering the nature of the memory organization that supports these remindings or how they are used to help in understanding. This paper will present a functional view of episodic reminding from the perspective of planning rather than understanding. Because of the clearer functionality of planning, a more straightforward view of the use of these remindings and the memory organization that supports them is possible. This paper will discuss three situations in which episodic memories can be used in planning: problem anticipation, plan construction and plan repair. The memory organization that each of these uses implies will also be discussed. This view of memory will be presented in terms of a case-based planner, CHEF, which makes extensive use of episodic remindings in constructing new plans.

The phenomena of reminding has been receiving quite a bit of attention in the past few years. Researchers have been looking at how previous experience can help in understanding and problem solving. As much of this work has shown, reminding is a complicated process. In this paper, we present requirements on a cognitive architecture that can promote and support these observations. The architectural requirements are those needed by a machine that can use experience to reason. To motivate these requirements, we present observations of reminding that are based on analyses of both actual remindings and hypothetical cases.

In this paper we present a case-study to demonstrate an application of concepts of knowledge representation from cognitive science and AI to problem solving in psychotherapeutic situations. In particular, a special type of frame, the so-called "Thematic Organization Point", or TOP, is used to characterize generic conflictive patterns of interaction, and to elucidate the meaning of a "psychotherapeutic interpretation". The concept of "failure-driven memory" is related to the process of evoking memories in patients. A belief systems analysis is used to explain why in some situations people are incapable of learning in spite of repetitive expectation failures. The underlying theory is summarized as a set of "Theorems". It is concluded that a cognitive science approach to therapeutic problem solving not only clarifies theoretical concepts but enables the derivation of powerful heuristics to be used by therapists in their practical work,

This paper discusses a method for representing thematic level structures, i.e. abstract plan/goal combinations. W e make the case that the processes for both recognition and construction of plans use the s a m e memory structures. In particular, w e are looking at the knowledge structures for recognizing and avoiding bad planning. The learning procedure w e describe starts by observing the bad planning behavior of narrative characters and combines old descriptions of planning errors to create new abstract structures. The learning method discussed is a one-trial, schema acquisition method, which is similar to DeJong's [DeJong 1983]. The method used involves taking schemas for planning situations that are found in an actual narrative situation, and using causal reasoning to construct a new schema which better characterizes the situation. This work is part of the M O R R I S project at U C L A [Dyer 1983a]. The planning situations are represented using Thematic Abstraction Units (TAUs) [Dyer 1983b.

In this paper, we present three examples of the development of legal concepts and rules analyze them in therms of AI learning methods, in particular, the candidate elimination algorithm. The three areas of legal doctrine are: (1) the doctrine of the liability of manufacturers to third party buyers; (2) the Fourth Amendment doctrine relating to stop-and-frisk searches; and (3) the doctrine of attractive nuisance. For each of the legal areas, we give a synopsis of the legal developments and show how it can be viewed from an Ai point of view.

A connectionist network has been used to simulate the solution, using a hill-climbing heuristic,of the DOG - > CA T puzzle (changing 1 letter at a time, generate a sequence of 3-letter wordsbeginning with DO G and ending with C A T ) and a simpler variant of the 8-tile puzzle, thedog-cat-mouse (DCM ) ^'uzzle devised by Klahr (1985). Distributed representations have beenused to represent the dilFcrent possible states of the puzzles. These states are learned by thenetwork and become local energy minima of the system. T o simulate the sequence of statescorresponding to a solution of the puzzle, the initial state of the network is set to the startstate, and the goal state is presented to the network as a continuous input. A sequence oi"states is generated by habituation, a short-term modification of the connection strengths whenever all the elements in the network are maximally or minimally activated, and by exploitingthe property that successive states comprising the solution are similar.

Stacks and trees are implemented as distributed activity patterns in a simulated neural network called BoltzCONS. The BoltzCONS architecture employs three ideas from connectionist symbol processing -- coarse coded distributed memories, pullout networks, and variable binding spaces, that first appeared together in Touretzky and Hinton's neural net production system interpreter. In BoltzCONS, a distributed memory is used to store triples of symbols that encode cons cells, the building blocks of linked lists. Stacks and trees can then be represented as list structures. A pullout network and several variable binding spaces provide the machinery for associative retrieval of cons cells, which is central to BoltzCONS' operation. Retrieval is performed via the Boltzmann Machine simulated annealing algorithm, with Hopfield's energy measure serving to assess the results. The network's ability to recognize shallow energy minima as failed retrievals makes it possible to traverse binary trees of unbounded depth without maintaining a control stack. The implications of this work for cognitive science and connectionism are discussed.

Fluent human sequential behavior, such as that observed in speech production, ischaracterized by a high degree of parallelbm, fiizzy boundaries, and insensitivity toperturbations, hi this paper, I consider a theoretical treatment of sequential behaviorwhich is based on data from speech production. A networi^ is discussed which tsessentially a sequential machine built out of connectionist components. The networkrelies on distributed representations and a hi(^ degree of parallelism at the level of thecomponent processing units. These properties lead to parallelism at the level at whichwhole output vectors arise, and constraints must be imposed to make the performanceof the network more sequential. The sequential tr^ectories that are realized by thenetwork have dynamic properties that are analogous to those observed in networkswith point attractors (Hopfield, 1982): learned tn^ectories generalize, and attractorssuch as limit cycles can arise.

As the human hand reaches out to grasp an object, it preshapes into a shape suitable for theanticipated interaction. As it gets close to the object, it encloses it. Behavioral studies haveshown interactions between grasping components and arm transport components. Biomechanicalstudies look at human hand postures, while neurophysiologists try to suggest how those posturesare formed and controlled. In order to evaluate such studies, models are needed which can beused to predict grasping behavior characteristics from a few basic object properties. By lookingat the action-perception cycle in primate hand movement, we hope to gain insight into neocorticalorganization, thus being able to suggest algorithms which could also be at work at all levels ofhuman intelligence.

A connectionist architecture is applied to the problem of 3-D visual representation. The Visual Perception System (VIPS) is organized as a flat, retinotopicallymapped array of 16K simple processors, each of which is driven by a coarselytuned binocular feature detector. By moving through its environment and observing how the visual field changes from state to state for various kinds of motion,VIPS learns to run internal simulations of 3-D visual experiences, e.g. mentalrotations of unfamiliar objects. Unlike traditional approaches to visual representation, VIPS learns to perform 3-D visual transformations purely from visualmotor experience, without actually constructing an explicit 3-D model of thevisual scene. Instead, the third dimension is represented implicitly in theknowledge as to how the pattern of activation on its fiat sheet of binocularlydriven processors will shift about as VIPS moves, or only imagines moving,through space. VIPS is argued to be more compatible with a variety ofphenomena from the psychology of 3-D perception than previous vision systems,particularly with respect to development, plasticity, and stability of perception, aswell as the analogical, linear-time mental rotation phenomena.

Dynamic properties of a model post-retinal, short-terra visualstore called a retinoid are presented. The retinoid modelgenerates specific predictions about visual perception in several variants of the seeing-more-than-is-there (SMTT) paradigm.Three experiments were conducted to test for the predictedeffects. Performance curves are presented and discussed. Anew visual illusion of motion is shown to follow from the properties of the retinoid model.

This paper presents the architecture of the discovery system SHUNYATA which models studies research in higher mathematics. SHUNYAT A analyzes mathematical proofs and product's concrpisand proof strategies which form the basis for the discovery of more difficult proofs in other mcichomatical theories. Its architecture avoids combinatorial explosions and does not, recijuire search str;uegies.The proof strategies contain two categories of predicates. A predicate of (he first caregoiy •^v-iirtsa small set of proof steps and the predicates of the second category evaluate partial proofs .uid decide which predicate of the first category should be apphed next. Thus, the proof strattgics includefeedback loops. A detailed example is given. It contains a simple proof in group theory, tlie .m-ilysisof this proof, and the discovery of a proof in lattice theory whose degree of diBiculty repremis r.hestate-of-the-art in automated theorem proving. The most important result of this work is tli".' discovery of a holistic logic based on the concept that cognitive structures arise from simple perceptions,evolve by reflection and finally contain their own evolution mechanisms.Keywords: Learning, knowledge acquisition, cognitive evolution, automated theorem proving.

Previous research has shown that stimulus values on a sensorycontinuum are perceived in a categorical manner by human subjects andby rhesus monkeys (Wilson, 1972; Streitfeld & Wilson, in press). Thatis, stimuli which are judged to belong to different perceptualcategories are discriminated more accurately than are stimuli whichare perceived as belonging to the same perceptual category. In theseexperiments, the category boundary was defined as the adaptation level(AL) established by the stimulus series presented to the subject.Wilson and DeBauche (1981) showed that resection of visual"association cortex" in monkey abolished categorical perception ofvisual features and they hypothesized that modality-specific neuralsubstrates that preserve the effects of stimulation provide aninternal referent which determines the manner in which given stimulusvalues are identified and discriminated. In the study described here,the effects of focal brain damage on categorical perception of threestimulus continua was examined in neurological patients. Processingof visual features as members of perceptual categories was doublydissociated from processing of haptic stimuli; posterior lesions inthe right hemisphere selectively impaired categorization of visualstimuli differing in length and orientation while anterior lesions inthe left hemisphere selectively impaired categorical perception ofweight. Implications for the neural dynamics of categorization arediscussed in the context of AL theory and principles of neuralorganization.

A framework for representing and reasoning with uncertain information is described.A network knowledge structure is used which makes the reasons for believing or not believing a proposition explicit. These reasons, or endorsements, are quantified by a measureof belief and certainty. Heuristics are integrated with the knowledge structure to collect, andevaluate the endorsements.

Conceptual structure in humans, pigeons, and monkeys was investigatedusing a multidimensional scaling procedure. Pigeons and monkeys wereinitially trained to discriminate between stimultaneously presented treeand nontree pictorial stimuli. Preference data was collected by insertingprobe trials In which the animals were forced to choose between two treestimuli. Analogous data for human subjects was collected by havingsubjects rate their preferences for the same stimuli. Tree preferencerelationships in the different datasets were obtained using the DEDICO Mprocedure. These analyses revealed striking interspecies differences inconceptual structure. The analysis of human tree preferences revealed a'whole vs. part' pattern in which stimulus preference was a function ofstimulus completeness. Pigeon tree preferences were qualitatively differentfrom human tree preferences, and also appeared to be less elaborate. Ingeneral, 'branchy' stimuli were preferred over 'leafy' stimuli, and a 'wholevs. part' pattern did not emerge. The data for monkeys also illustrated apreference for 'branchy' structures over 'leafy' structures. Individualdifferences between monkey preferences were also revealed, and were foundto be related to performance on the initial discrimination task. Monkeysthat had a well-defined tree preference pattern learned this task fasterthan did monkeys with a less defined structure. The results of all of theanalyses demonstrated interspecies differences in tree concepts, and suggested the possibility that these may be related to different functional experiences or requirements.

A computational model of a motor program generator (MPG) forhorizontal pursuit eye movements (PEM) is proposed. The MPG modelconsists of two neural networks (velocity maps). Neurons arearranged in a single circular layer with lattice structure andconnected only to their immediate neighbors. During PEM one ofthe two maps always features an activity peak (AP) which travelswith constant velocity from one neuron to the next. A memory traceof the most recent portion of the trajectory is created by meansof a temporary increase of the interneuronal connectivity strengthbetween previously activated neurons. This novel MPG model may beuseful for designing parallel processors for motor control ofrobots.

Machine learning methods for connectionist models usuallyoperate by attaching weights to a prespecified network so that a certainfunctionality is achieved. This is the classical credit assignment problem.This paper explores a constructive approach to connectionist learningwhere both a network and weights must be generated. It is argued that thisis an easier problem to solve and is sufficient for many applications sincenetwork topology is usually not as important as functionality.Three algorithms are presented for constructing networks from trainingexamples. A s cells are added and iterations are made, each method produces a network having optimal expected behavior (i.e. it correctly classifiesthe maximu m number of training examples possible) with arbitrarily highprobability p < 1.Learning speed for these algorithms is currently being investigated.

How is affect communicated in language? Natural languages contain a large number oftechniques for injecting affect into text, both explicitly and implicitly. This paper discussessome techniques that speakers use to slant their text, and describes what is required for acomputer program to generate differently slanted versions of a single underlying representation.

Two experiments tested the predictions of an integrative buffer model of visual processing, regarding the illusory conjunction of components of rapidly presented displays. Color pictures of objects were presented at a rate of 9/s, in the same spatial location. Experiment 1 used a modified report procedure to test the hypothesis that Stroop-like response competition, during naming, not a perceptual error, resulted in the high confidence "illusory conjunctions" reported in previous research. Subjects were provided with the naem of a picture in advance and reported "yes" or"no" to indicate if that picture was the one in the frame. Contrary to the response competition hypothesis, high confidence errors occured frequently under these conditions. Experiment 2 tested the hypothesis that the direction of migration (preceding or following picture) is the result of a difference in the sequential allocation of attention to the frame first or to its "host" picture first on different trials. As predicted by the integrative buffer model, subjects were faster in detecting the frame when they confidently reported it around the preceding picture than around the followuing picture in the sequence, and reaction times associated with correct reports fell between the two.

We investigated the effects of spelling knowledge on the representationof spoken language. In the experiment, subjects first saw the written form ofa nonsense word, then heard it and judged the number of syllables. For theidentical acoustic tokens, the number of judged syllables varied with theaccoitpanying spelling. The effect of spelling was stronger for one syllablepronunciations than for two syllable pronunciations. The results arediscussed in relation to the role of spelling knowledge in listeners'representations of phonology.

We describe a system that learns new schemas by modifying old ones, in order to understand anomalous events in stories that it reads. W e discuss how these schemas (called Explanation Patterns[Schank 86]) are structured in order to make them modifiable, and how the understanding processapplies and modifies them. This model bridges the gap between two previous models of understanding,which were based on either application of prestored schemas or understanding-time inference chaining.By employing modifiable schemzis, our model is more flexible than the former and more efficient thanthe latter.

We present a mechanism for remembering explanations and re-using them to explain newepisodes. This task requires a representation scheme for explanations, a dynamically organizedmemory, and a means of modifying old explanations to fit new facts. In this paper we focus onmemory organization. W e describe strategies for indexing and retrieving explanations, for usingcausal knowledge to select relevant features of episodes and for guiding generalization. W e discusswork in progress on a computer implementation of this model.

Metaphor is a problem for natural language knowledge acquisition systems. Experts will make utterances based upon domain metaphors whichthe acquisition system ma y not possess. A n approach is presented whichuses knowledge about previously understood metaphors to process new uses.This approach is contrasted with several formal proposals for metaphorunderstanding which do not use explicit knowledge about metaphors. Asystem for representing metaphorical knowledge, as part of a generalknowledge representation language, has been built. A knowledge acquisition system, UCTeacher, is described which can process newly encounteredmetaphors using knowledge of the domain and explicit knowledge abouthow similar metaphors have been used before. A detailed example from thesystem is presented.

The design and architecture of two user-controlled, computer-based problemanalysis environments in classical mechanics are discussed. In theexpert-like environment, the user analyzes problems according to ahierarchical concept schema consistent with how experts analyze novel problemsin physics. In the second environment, the user searches a large equationdata-base utilizing novice-like, surface feature keywords in order to locatethe appropriate equation(s) to use in solving a problem. Cognitive andpedagogical implications of the research are discussed.

Several dynamical systems have been previously pTX>posed to give a neural-Iike (i.e. connectionist)description of category formation. These typically either involve supervised training (as in Sutton &.Barto, 1981; Reilly et al., 1982) or identify dense regions ("clusters') in the stimulus distribution asnatural categories (Amari & Takeuchi, 1978; Rumelhart & Zipser, 198S). By combining two existingconnectionist-type learning procedures, one supervised and one unsupervised, a hybrid 'self-supervisedleamingf (SSL) mechanism for concept and category learning has been developed. Each unit in thenetwork comes to represent some concept of the order of complexity of a single word; the activity ofthe unit signals the contribution of its associated concept to the current mental state. A crucialassumption of this i^proach is that every concept unit (C-unit) receives inputs from two or moreinformation streams. The self-supervised learning process is governed by a data-driven dynamical rulewhich results in a two-stage learning process. In the first stage, a C-unit becomes selectively responsiveto a particular pattern •"'• from one of the information streams, ignoring all other patterns in thatstream. This is followed by an associative stage in which the unit develops graded response propertiesto stimulus patterns incident from the other information stream(s). The trigger feature thus becomes akind of prototype for the concept to be formed by the C-unit. Populations of C-units display interesting representational properties; these are seen to have attributes of both local and distributed representations.

Context effects on lexical decisions were anatomised bymanipulating lexical relatedness in syntactic andasyntactic sequences. In a Syntactic condition, relatedor unrelated word-pairs were embedded in simplesequences (e.g., a truck or a CAR/FLOOR). In aScrambled condition, two inapposite function words weresubstituted between the related and unrelated nouns(e.g., the truck that before CAR/FLOOR). The phraseswere presented serially and subjects made lexicaldecisions to their terminal elements. Substantialrelatedness effects were found only in syntacticsequences, whether presentation rate was slow or whetherit exceeded the rate of normal reading. The syntacticrelatedness effect was shown to consist, in equalproportions, of facilitation of related words andinhibition of unrelated words. These results argueagainst a role for intralexical priming in on-linereading. They point up the roles of syntacticconnectedness and of the current interpretation even invery rudimentary contexts.

Although mental schemata are central to many contemporary learning theories, theprecise relationship between such schemata and specific types of learning remains vague.This paper describes an analysis of subjects learning basic computer skills when presentedwith four different kinds of elaborations that should influence subjects' schemata: (I) asimple description with no model: (2) a redundant elaborated text also with no explicitmodel: (3) a functional model: (4) a descriptive analogy. Subjects were tested onprocedures, general comman d concepts, and system questions. Models and analogies wereshown to improve initial performance on all types of questions. However only systemquestions showed this advantage after a delay. It is argued that the utility of building amental model through elaboration depends on the specific tasks that are analyzed.

We propose that three qualitatively different strategies help focus attention during problem solving.The first strategy is to apply operators that will lead to definite progress toward the goal. Attention will befocussed by this strategy as long as some operator in this class is applicable, Whe n clear progress can not beachieved, the problem solver must decide how best to proceed. It then invokes the second strategy to selectoperators that preserve important characteristics of the current problem. These operators are likely to keepthe problem solver from diverging sharply from the goal while possibly enabling the application of operatorsby the first strategy. Whe n the problem solver can follow neither of the first two strategies, it invokes thethird strategy of arbitrarily applying legal operators. W e see the second strategy as an essential differencebetween novice and expert problem solvers. It is easy to recognize definite progress to^*'ard a goal and it iseasy to recall which operators can be legally applied. Expertise involves knowing which characteristics of asituation should be preserved (or created) when no way to definitely progress toward the goal is known. Thisthree-strategy theory has been implemented and tested in a system that performs mathematical calculationsin the course of solving physics problems. W e describe a number of mathematical calculation operators usedunder each strategy

This article contributes to the theory of network learning procedures by identifyingand analyzing two problems with the backpropagation procedure of Rumelhart, Hinton,and Williams (1985) that may slow its learning. Both problems are due to backpropagation's being a gradient- or steepest-descent method in the weight space of the network. Thefirst problem is that steepest descent is a particularly poor descent procedure for surfacescontaining ravines—places which curve more sharply in some directions than others—andsuch ravines are common and pronounced in performance surfaces arising from networks.The second problem is that steepest descent results in a high level of interference betweenlearning with different patterns, because those units that have so far been found most useful are also those most likely to be changed to handle new patterns. The same problemsprobably also arise with the Boltzmann machine learning procedure (Ackley, Hinton andSejnowski, 1985) and with reinforcement learning procedures (Barto and Anderson, 1985),as these are also steepest-descent procedures. Finally, some directions in which to lookfor improvements to backpropagation based on alternative descent procedures are brieflyconsidered.

A Neuronal model is described that can parse, learn, and recognize objects in a complex visual environment. A comptuer simulation of the model network was tested witha variety of scenes and exhibits competent performance

In this contribution we investigate the mental representation of spatial information. In a previous paper (Wender & Wagener, 1985) we reported results that suppoerted the idea of spatial information being mentally represented in an analogue format. THis paper extends the previous results. Furthermore a simulation model is developed that descries the data from two different experimental tasks, a priming technique and a sentence verification test. These experimental tasks gave results that in part look contradictory. The model discussed is an attempt to account for this contradiction.

The back-propagation learning algorithm (Rumelhart, Hintoa, & Williams, 1986) for connectionist networks works by adjusting the weights along the negative of the gradient in weight space of astandard error measure. The back-propagation technique is simply an efficient and entirely local means of computing this gradient. Using what is essentially the same back-propagation scheme, onemay instead compute the gradient of this error measure in the space of input activation vectors; thisgives rise to an algorithm for inverting the mapping performed by a network with specified weights.In this case the error is propagated back to the input units and it is the activations of these units —rather than the values of the weights in the network — that are adjusted so that a specified outputpattern is evoked. This technique is illustrated here with a small network which is a much simplifiedversion of the NETtalk text-to-specch network studied by Sejnowski and Rosenburg (1986). The ideais to run this network backward so that it attempts to spell words based on their phonetic representations. This example further illustrates the use of this technique in a sequential interpretation settingin which phonemes are presented to the system one at a time and the system must refine its previousguess at the correct spelling as each new phoneme is presented.

This study is on the extended line of Grimshaw 1979,which explains the complement selection in Japanese. Byextending Grimshaw's analysis that the combination ofpredicates and their complements are explicable by imposing well-formedness conditions on two differentlevels of representation: one at the syntactic level;the other at the semantic level, the analysis givenhere which utilizes two semantic restrictive featuresunder the semantic feature: [+presupposition] and[±factive], is able to explain the anomalies concerningthe complementizer selection in Kuno 1973.

This paper describes an implemented tutoring system, designedto show various ways of converting a given meaning structureinto its corresponding surface expression. The system is meantto be a teaching tool for students who learn French as a foreignlanguage.Vihile showing various ways of converting a ijiven ineaning structure into its corresponding surface expression, the system helpsnot only to discover WHAT data to process but also MOW this information processing should take place. In other words, we are concerneiJ wii't Hrficiency in verbal planning (flexibility and economy of performance).Recognizing that the same result can be obtained by various methods, the student should find out which one is best suited tothe circumstances (what is known, task demands etc). Informational stato:5, lienoe tlie processor's needs, may vary to a great extent, as may his STRATEGIES or cognitive styles. In consequence,in order to become an efficient processor, the student has toacquire not only STRUCTURAL or RULE-KNOWLEDGE but also PROCEDURALKNOWLEDGE (skill).With this in mind we have designed three modules in order tofoster a reflective, experimental attitude in the learner, helping hln to discover insightfully the most efficient strategy.