Proceedings of the Annual Meeting of the Cognitive Science Society

We describe a recurrent neural network model of rate effects on the syllable-initial voicing distinction, specified by voiceonset-time (VOT). The stimuli were stylized /bi/ and /pi/ syllables covarying in VOT and syllable duration. Network performance revealed a systematic rate effect: as syllable duration increases, the category boundary moves toward longer VOT values, mirroring human performance. Two factors underlie this effect: the range of training stimuli with each VOT and syllable duration, and their frequency of occurrence. The latter influence was particularly strong, consistent with exemplar-based accounts of human category formation.

We propose that causal attribution involves constructing a coherent story using mechanism information (i.e., the processes underlying the relationship between the cause and the effect). This processing account can explain both the conjunction effect (i.e., conjunctive explanations being rated more probable than their components) and the discounting effect (i.e., the effect of one cause being discounted when another cause is already known to be true). In the current experiment, both effects occurred with mechanism-based explanations but not with covariationbased explanations in which the cause-effect relationship was phrased in terms of covariations without referring to mechanisms. We discuss why the current results pose difficulties for previous attribution models in Psychology and Artificial Intelligence.

A cognitive account of propositional reasoning must consider both the representation of the propositions (premises and states of affairs) and the context in which the propositions are used. This paper is concerned with reasoning processes involving three different connectives (conjunctive, conditional and disjunctive connectives) in three different tasks (accomplishing a request for action expressed by a premise, judging a state of affairs as true or false with respect to a premise, drawing an inference from two premises). Our claim is that the ability to reason with connectives is explained in terms of construction and manipulation of mental models. We present a computer model that takes as input the modelistic representations of the premises and the speciHc state of affairs, compares such models and gives rise to a series of model manipulations in order to produce a result, i.e. an action, a judgement or an inference. A computer program reproduces the performances of subjects of different age groups, predicting both correct and erroneous inferences.

An unprecedented combination of simulative and metaphor-based reasoning about beliefs is achieved in an AI system, ATT-Meta. Much mundane discourse about beliefs uses conceptusJ metaphors (e.g., MINDAS CONTAINER ) productively, and ATT - Meta's metaphor-based reasoning accordingly leads to crucial discourse comprehension decisions. ATT-Meta's non-metaphorical mode of belief reasoning includes simulative reasoning (SR). In ATT-Meta, metaphor-based reasoning can block and otherwise influence the course of SR. Also, ATT-Meta can nest SR and metaphorbased reasoning within themselves and each other. As well as currently allowing ATT-Meta to simulatively reason about beliefs about beliefs ..., the nesting will in the near future allow the system to handle chained metaphors, ascribe its own metaphor-based reasoning to other agents, and apply simulative reasoning to purely metaphorical agents.

Populations of simple recurrent neural networks were subject to simulations of evolution where the selection criterion was the ability of a network to learn to recognize strings from context free grammars. After a number of generations, networks emerged that use the activation values of the units feeding their recurrent connections to represent the depth of embedding in a string. Networks inherited innate biases to accurately learn members of a class of related context-free grammars, and, while learning, passed through periods during which exposure to spurious input interfered with their subsequent ability to learn a grammar.

Research in case-based design has demonstrated some capability to retrieve relevant designs and to adapt them automatically to satisfy new design constraints. However, some domains are less amenable to automated adaptation, particularly when the cases are very complex and when relationships among the design components are difficult to express formally. The design of interactive learning environments is one such domain. W e describe a case-based approach to instructional software design which utilizes interactive, model-driven case adaptation. Our model for computer-based instruction is Goal-Based Scenarios. W e describe a tool, Goal-Based Scenario Builder, which supports interactive adaptation of instructional software using the model, and illustrate its use in adapting an example case of a successful instructional software program, Sickle Cell Counselor.

Individual differences in collaborative explanations during learning were analyzed to detennine effects on problem solving. Twenty-five university students with no prior programming experience worked through a sequence of programming lessons. For the Target lesson, subjects studied instructional texts and examples in either mixed performance-level dyads (collaborative dyad group) or individually (individual group) prior to individual programming activities. The collaborative dyad subjects were divided into equal sized groups of high-benefit and low-benefit dyad subjects based on Target lesson programming performance. Betweengroup analyses of the characteristics of the explanations generated by high-benefit and lowbenefit dyad subjects were investigated, including (a) explanation and metacognitive strategies, (b) content of elaborations, and (c) manner of generating elaborations. High-benefit dyad subjects were found to generate both a higher quantity and higher quality of elaborations. These results are compared to findings from prior research

The selective vulnerability of morphology in agrammatic aphasia is often interpreted as evidence that closed-class items reside in a particular part of the brain (i.e., Broca's area); thus, damage to a part of the language processor maps onto behavior in a transparent fashion. W e propose that the selective vulnerability of grammatical morphemes in receptive processing may be the result of decrements in overall processing capacity, and not the result of a selective lesion. We demonstrate agrammatic profiles in healthy adults who have their processing capacity diminished by engaging in a secondary task during testing. Our results suggest that this selective profile does not necessarily indicate the existence of a distinct sub-system specialized for the implicated aspects of syntax, but rather may be due to the vulnerability of these forms in the face of global resource diminution, at least in grammaticality judgment.

In many domains, the content of a problem (i.e., its surface cover story) provides useful clues as to the type of problem it is and its solution. Three experiments examined this role of problem content on the problem categorization and solution of algebra word problems with experienced subjects, by manipulating only the content of the problems. When a problem's content was highly correlated with its deep structure (e.g., a content of cars driving for a distance-timerate problem), people were able to categorize the problem after seeing a smaller portion of it compared to a baseline with contents uncorrelated to the problem deep structure. In addition, for more complex problems in which irrelevant information had been added, problem solving performance was higher and people showed greater sensitivity to the relevance of the information. When a problem's content suggested a different (inappropriate) type of problem, people required a greater part of the problem to categorize it and were slower and less accurate at solving the problem. These results suggest that content may be influential even for experienced problem solvers.

Kripke (1972) and Putnam (1975a; 1975b) have argued forcefully for the philosophical view of word meaning known as rigid designation. While certain psychological studies have appeared to offer this view support (Keil, 1986; Rips, 1989), we argue that these have not provided an exhaustive evaluation. In particular, the original discussions of Kripke and Putnam reveal that their view rests on an explicit appeal to intuition concerning word use in a range of different scenarios. The study reported here investigates word use under three such types of scenarios, for a variety of natural kind terms, by investigating subjects' judgements of truth or falsity for a range of statement types. W e argue that the results obtained indicate that the intuition on which rigid designation rests is not one which is generally true of agents' language use. Further, we obtam patterns of apparent contradiction which appear strictly inconsistent with rigid designation and which require an account of word meaning which allows that the sense of words may vary systematically with context (Franks & Braisby, 1990).

Most philosophers of science now believe that scientific data are theory laden, i.e., the evaluation of data is influenced by prior theoretical beliefs. Although there is historical and psychological evidence that is consistent with the theory-laden position, experimental evidence is needed to directly test whether prior beliefs influence the evaluation of scientific data. In a fully counterbalanced design, one group of subjects received evidence that dinosaurs were cold-blooded, and another group of subjects received evidence that dinosaurs were warm-blooded. The subjects reported a strong belief in whichever theory they had read about. Then subjects were presented with a piece of data that supported one theory and contradicted the other theory. The identical piece of data was rated as more believable when it was consistent with the subject's theory than when it was inconsistent. These results provide clear support for the position that scientific data are theory laden.

Some of Kant's ideas about the mind have had a huge influence on cognitive science, in particular his view that sensory input has to be worked up using concepts or concept-like states and his conception of the mind as a system of cognitive fubnctions. We explore these influences in the first part of the paper. Other ideas of Kant's about the mind have not been assimilated into cognitive science, including important ideas about processes of synthesis, mental unity, and consciousness and self-consciousness. They are the topic of the second part of the paper.

Connectionist simulations of children's acquisition of velocity (v), time (t), and distance (d) concepts were conducted using a generative algorithm, cascadecorrelation (Fahlman & Lebiere, 1990). Diagnosis of network rules were consistent with the developmental course of children's concepts (Wilkening. 1981, 1982) and predicted some new stages as well. Networks integrated the defining dimensions of the concepts first by identity rules (e.g.. v = d), then additive rules (e.g., v = d-t), and finally multiplicative rules (e.g., v = d/t). Psychological effects of differential memory demands were also simulated. It is argued that cascade-correlation implements an explicit mechanism of developmental change involving incremental learning and qualitative increases in representational power.

We present a new connectionist model of human spelling and investigate some of its properties. Although based on Sejnowski & Rosenberg's (1987) NETtalk model of reading, it requires no pre-processing of the training data to align the phonemes and letters. The model achieves 100% performance on the training data (2837 monosyllabic words including many irregular words) and has a generalization performance of about 89%. Under appropriate conditions it exhibits symptoms similar to developmental surface dyslexia and acquired surface dysgraphia. However, its inability to account for phonological dysgraphia and lexical decision leads us to believe that it is a promising candidate for the rule based part of a dual route model but not a complete model of spelling on its own.

We use hierarchical cluster analysis, principal component analysis, multi-dimensional scaling and discriminant analysis to investigate the internal representations learnt by a recent connectionist model of reading aloud. The learning trajectories of these representations may help us understand reading development in children and the results of naming latency experiments in adults. Studying the effects of network damage on these representations seems to provide insight into the mechanisms underlying acquired surface dyslexia. The discussion of the various techniques used may also prove useful in analysing the functioning of other connectionist systems.

We implement a constraint satisfaction connectionist style model that accounts for data from three psycholinguistics experiments investigating the gardenpaih effect with reduced relative constructions. Normative data was collected on the stimuli used in experiments by Burgess and Tanenhaus (1992) and Ferreira and Clifton (1986) and this data served as the input for the simulation. W e have demonstrated with this set of simulations that a plausible theoretical framework for a range of these results is a hierarchical connectionist network which is sensitive to a number of constraints inherent in the input stimuli. The model accounts for the top-down effect of context, the contribution of the bottom-up morphological frequency asymmetry of the verb, and the probabilistic nature of the disambiguating preposition. These effects are sensitive to the timecourse of processing as well. The pattern of results from the psycholinguistic data suggest that syntactic processing is a confluence of multiple constraints that represent both bouomup and top-down influences in processing. These results are incompatible with a deterministic parsing model. The hierarchical connectionist style model presented in this paper is sensitive to the range of constraints discussed above and is offered as a more adaptive theoretical model that can capture the domain of effects found in the literature encompassing local syntactic ambiguity resolution.

Subjects were presented with strongly past-participle biased sentences such as, Tlie portrait sketched by the tree was very beautiful, in a self-paced reading time task. Sentences were displayed two words at a lime, (e.g.. The portrait /sketched by ...) so that the verb and the disambiguating preposition were read together. In Experiment 1, a set of materials constructed to minimize the past-tense bias with an inanimate N P was compared with a less constraining set of sentences. The syntactic gardenpalh usually associated with the reducedrelative construction was not present with the more constraining materials, but was with the less constraining N P sentences. In Experiment 2. using only the more constraining materials, preposition lengdi was manipulated so that subjects read sentences with both short (i.e., by) and long (i.e., underneath) prepositions. No syntactic gardenpaths occurred with sentences with the past-participle bias and short prepositions; however, when the same sentences were read with the long prepositions, the syntactic gardenpath was present This result is inconsistent with a deterministic parser. W e expand on our previous proposals that the parser must be able to take into account both semantic and verb-form information, as well as, the amount of disambiguating information (in the form of a preposition) that can be integrated with the ambiguous verb.

ACME and Copycat have been viewed as competing models of analogy making. Mitchell (1993) makes three major criticisms of ACM E in arguing for Copycat's superiority: that because ACME considers all syntactically possible mappings it is psychologically implausible and computationally infeasible; that its representations are rigid and hand-tailored for each problem; and that ACME's representations are scmantically empty. To evaluate these criticisms we applied ACME to simulating problems in the only domain addressed by Copycat, letter-string analogies such as, "If abc is changed into abd, how would you change kji in the same way?" Using representations that include only knowledge available to Copycat, ACME generated the most common solutions that people and Copycat produce. In addition, ACME was able to generate some solutions produced by people but that are impossible for Copycat, demonstrating that in some respects ACME is a more flexible analogical reasoner than is Copycat. These simulations answer each of Mitchell's criticisms of ACME . ACME can incorporate domain-relevant knowledge to allow a principled reduction in the number of mappings considered; it can generate novel representations based on its domain-general constraints; and it can incorporate semantic content into its representations. In addition, ACME has the advantage of being applicable to many different domains.

The age of a case (in the CBR sense) is the amount of time that has elapsed between the time that the case originally occurred and the time of the current reasoning activity. People engaged in plausible reasoning tasks will, under appropriate circumstances, use the age of retrieved prior cases to filter and discard them, or to select among alternatives by their recency. This paper examines how the age of a case (and its spatial analog) are used by people in plausible and case-based reasoning tasks. I will argue that (1) The age of a retrieved case is an important factor in case relevance judgments for certain kinds of inferences. (2) When case age is relevant, more recent cases are usually, but not always, preferred to older ones (the all other things being equal" caveat). Finally, I will argue that, somewhat surprisingly, (3) case age cannot be used as an index into memory given some commonly held assumptions about the nature of the retrieval process because it varies with the time of retrieval. This limits its use to post-retrieval processes, such as the filtering of already retrieved cases.

How can misreported information be effectively corrected? Wilkes and Leatherbarrow (1988) found that people relied upon invalidated information to answer questions despite their awareness of its inaccuracy, a phenomenon called the "continued influence effect" (Johnson & Seifert, in press). But corrections in which an assertion is made and then denied (e.g., "X is true ... actually, X is untrue") ma y violate important conversational assumptions. Grice (1967/1989) and others have argued that people expect speakers to offer only information that is both truthful and conversationally relevant; thus, people may seek interpretations for corrections that will incorporate both the literal meaning and the conversational implications of the contradictory statements. Our hypothesis was that corrections would be more successful when they explained why the original information was asserted. An empirical study showed that corrections that accounted for conversational implications (e.g., "X, which bad originally been believed because of Y, is actually untrue") could more effectively reduce the continued use of discredited information. Additionally, the results show that reiterating the literal content of a correction ma y actually be perceived as implying that the correction statement should be disbelieved. Since the conversational implications of corrections critically shape comprehension, their examination is crucial in domains (such as courtrooms, newspapers, and classrooms) where informational updates frequently occur.

Hypothetical reasoning - thinking about what might happen in the future or what might have happened in the past - enables us to go beyond factual reality. W e suggest that human reasoners construct a more explicit mental representation of hypothetical conditionals, such as. If Linda were in Dublin then Cathy would be in Galway, than of factual conditionals, such as, if Linda is in Dublin then Cathy is in Galway. When people think about the factual conditional, they keep in mind the affirmative situation -- Linda is in DubUn, Cathy is in Galway, and they maintain only an implicit awareness that there may be alternatives to this situation. In contrast, when they think about the hypothetical conditional, they keep in mind not only the affirmative situation, but also the presupposed negative one (Linda is not in Dublin, Cathy is not in Galway). The postulated differences in mental representations lead us to expect differences in the frequency of inferences that people make from the two sorts of conditionals, and we report the results of an experiment that corroborates this prediction. The psychological data have implications for philosophical and linguistic accounts of counterfactual conditionals, and for artificial intelligence programs designed to reason hypothetically.

Behavior analysis has recently developed a new paradigm for the study of categorization and language based on the mathematical notion of equivalence. Inspired by this paradigm, this paper presents a defmitional framework that could be relevant for several of the phenomena under study in Cognitive Science. First, categories are viewed as classes of functional equivalence. By doing so, results from behavior analysis and cognitive psychology seem to converge towards an experience-based interpretation of category basicness. Second, conditional equivalence is proposed as the basis for symbol-meaning and symbolsymbol relationships. Transfer of function through conditional links is suggested as the mechanism of connection between language and other aspects of cognition. The adoption and extension of these functionalist formalisms provides us with significant methodological, conceptual and even empirical advantages.

The use of transitional probabilities between phonetic segments as a cue for segmenting words from English speech is investigated. W e develop a series of class-based n-gram and feature-based neural network models that enable us to quantify the contribution of low-level statistics to word boundary prediction. Training data for our models is representative of genuine conversational speech: a phonological transcription of the London-Lund corpus. These simple models can be purely bottom-up and hence valid bootstrapping models of infant development. W e go on to demonstrate how the boostrapping models mimic the Metrical Segmentation Strategy of Cutler and Norris (1988), and we discuss the implications of this result.

Infants face the difficult problem of segmenting continuous speech into words without the benefit of a fully developed lexicon. Several information sources in speech—prosody, semantic correlations, phonotactics, and so on—might help infants solve this problem. Research to date has focused on determining to which of these information sources infants might be sensitive, but little work has been done to determine the usefulness of each source. The computer simulations reported here are a first attempt to measure the usefulness of distributional and phonotactic information in adult- and children directed speech. The simulations hypothesize segmentations of speech into words; the best segmentation hypothesis is selected using the Minimum Description Length paradigm. Our results indicate that while there is some useful information in both phoneme distributions and phonotactic rules, the combination of both sources is most useful. Further, this combination of information sources is more useful for segmenting childdirected speech than adult-directed speech. The implications of these results for theories of lexical acquisition are discussed.

This paper describes an implemented system that generates spoken dialogue, including speech, intonation, and gesture, using two copies of an identical program that differ only in knowledge of the world and which must cooperate to accomplish a goal. The output of the dialogue generation is used to drive a three-dimensional interactive animated model - two graphic figures on a computer screen who speak and gesture according to the rules of the system. The system is based upon a formal, predictive and explanatory theory of the gesture-speech relationship. A felicitous outcome is a working system to realize autonomous animated conversational agents for virtual reality and other purposes, and a tool for investigating the relationship between speech and gesture.

It is hypothesized that labels in examples help learners group a set of steps and to try to explain why those steps belong together. The result of these grouping and selfexplanation processes might be the formation of a subgoal. It is conjectured that the meaningfulness of the label itself might not be critical in order for the grouping and self-explanation processes to occur. This conjecture is supported in an experiment in which subjects studying examples in probability that had steps labeled transferred to novel problems more successfully than subjects whose examples did not contain labels. Furthermore, subjects who saw less meaningful labels transferred as successfully as subjects studying examples with more meaningful labels. Thus, it appears that the meaningfulness of the label does not seem to affect subgoal formation as much as the presence of a label. This result supports the interpretation that subgoal learning is affected by labels and that labels produce this benefit by helping learners group the steps into a purposeful unit, perhaps through a self-explanation process.

Logics of belief are usually either quite complex, unintuitive, make overly idealistic assumptions, or all of the above, because they have to cope with the unusual characteristics of the belief operator (relation, predicate). Some of these problematic characteristics are referential opacity, the possible falsehood of objects of belief, belief recursion, identification of referents from outside of the belief operator in quantification contexts, etc. The difficulties faced by traditional logical treatments seem to stem mainly from the fact that an essentially subjective, intensional phenomenon gets analyzed from an objective, outside observer's point of view in an extensional, logical framework. As an alternative, we propose a subjective, intensional logic SL, which takes seriously the usual characterization of belief as a propositional attitude, that is, in SL belief is treated as a relation between an agent and a proposition (an intensional object). As results we gain technical simplicity and a simple, intuitive semantics for belief sentences.

Law Encoding Diagrams, LEDs, are knowledge representations that correctly encode systems of one or more laws using the geometric and/or the topological strucnire of diagrams. In an instructional role, LEDs aim to focus learning on the formal relations defined by the correct laws, whilst using diagrammatic representations to aid comprehension. LEDs can be viewed as intermediate representations that aim to bridge the conceptual gulf between abstract laws and the behaviour of phenomena. It is anticipated LEDs will be adopted as key models in the foundation of expertise. This paper describes an investigation in which LEDs for momentum and energy conservation were used for instruction. The LEDs were implemented in a computer based discovery learning environment and the subjects given only minimal instruction on their use in problem solving. However, half the subjects used the LEDs for successful post-test solutions of different classes of problem and exhibited strategies that were expert-like, in marked contrast to their novice-like pre-test performance.

Philosophers have disagreed about whether theories that make successful predictions are more believable than theories that merely explain data that have aheady been discovered. Predictivists believe that theories that make successful predictions have an edge over theories that offer only retrospective explanations of the same data. Nonpredictivists maintain that whether a theory predicts data or explains data retrospectively is irrelevant to the believability of the theory. The purpose of this paper is to report on three psychological experiments designed to determine whether undergraduates behave as predictivists or nonpredictivists when they evaluate theories. Results indicate that subjects behaved as nonpredictivists when one theory predicted a body of data and a second theory was devised later to explain the same data retrospectively. However, subjects behaved as predictivists in the situation in which a theory retreated in the face of anomalous data by adding an auxiliary hypothesis; for instance, theories that predicted data by adding the necessary auxiliary hypotheses before the data came in were more believable than theories that added the auxiliary hypothesis in reaction to the data. These results suggest that cognitive models of theory choice that assume that people are nonpredictivists may require modification.

This paper analyzes two converging views of the architecture of intuition. A. diSessa and L. Talmy, working independently in different fields (physics education and linguistics), have formulated strikingly similar theories of intuition. Both view people's intuitions about forces as simple pieces of knowledge organized heterarchically. However, Talmy's force dynamic patterns have more system-wide structure than diSessa's phenomenological primitives. Using these primitives, people generate common sense explanations for a wide variety of situations. Moreover, people may build upon these intuitions while studying formal disciplines such as physics. However, several primitives directly conflict with physics concepts and may account for resilient misconceptions. Finally, intuitions may also provide the basis for understanding social and psychological phenomena.

Cognitive grammar provides an analytic framework in which the semantic value of linguistic expressions is characterized relative to domains of presupposed knowledge. Cognitive metaphor theory holds that metaphorical language involves a mapping of conceptual structure from a source domain to a target domain. Containers are one such pervasive structure. This investigation proposes a detailed representation for the domain CONTAINER and applies it in the analysis of metaphorical expressions mapping CONTAINER onto target domains ARGUMENT and linguistic expression. Each source domain word is analyzed with respect to which aspects of the CONTAINER domain structure it refers, and whether it refers to a 2D or 3D bounded region. The pattern of aspects mapped suggest that spatial containment, content, and material container object comprise major aspects of the 3D CONTAINER domain. The target domains are demonstrated to be structured according this container organization. The results demonstrate that cognitive semantic analysis can reveal specific structures of commonsense knowledge which are prerequisite for language use.

In this paper, we provide a taxonomy of the processes which people use to generate questions for a type of interviewing task. Specifically, we analyze "story acquisition interviews" in which the interviewer is a knowledge engineer who asks questions of a domain expert to acquire material for a conversational hypermedia system. Such interviews have proven to be surprisingly difficult to conduct successfully. W e have identified a number of "local" strategies which successful interviewers use to develop coherent, interesting sequences of questions and we have positioned these strategies within a model which describes the global interviewing process. This descriptive model is an initial step towards a methodology prescribing how to perform these interviews effectively.

This paper discusses evidence from thinking aloud case studies indicating that part of the knowledge used by expert problem solvers consists of concrete physical intuitions rather than abstract verbal principles or equations. One purpose of the paper is to provide empirical documentation of behaviors such as spontaneous references to using intuition, depictive hand motions, and dynamic imagery reports. Although the role of imagery in lower level tasks is becoming more accepted, we currently lack sufficient empirical evidence for its use in higher level thinking. In order to account for cases where subjects appear to be "running a simulation" of an event on the basis of a physical intuition, a model is presented in which a somewhat general and permanent perceptual motor schema controls a more specific and temporary image of a situation. This process is termed "imagistic simulation". The imagery can be kinesthetic as well as visual, and dynamic rather than static, suggesting the involvement of the motor system. Although rules for making inferences from networks of causal relations have been studied, we lack models which analyze the nature of mental simulations underlying a single causal relationship. Such physical intuitions and simulations may provide basic building blocks for constructing visualizable models in science.

We present a new variant of a simple recunent network to model auditory word recognition in continuous speech and address the issue of lexical segmentation. Simulations ba-sed on small word sets show that the system provides a nearoptimal solution to the opposite constraints of speed, which requires that lexical processing he immediate, and reliability, which imposes that identification decisions postponed until unambiguous information is available. Contraiy to an oftenheard statement, the simulations show that the existence of embedded words is not incompatible with the notion of continuous on-line lexical processing.

This paper evaluates the student modeling procedure in the ACT Programming Tutor (APT). APT is a practice environment that provides assistance to students as they write short programs. The tutor is constructed around a set of several hundred programming rules called the ideal student model, that allows the program to solve exercises along with the student. As the student works the tutor maintains an estimate of the probability that the student has learned the rules in the ideal model, in a process we call knowledge tracing. The cognitive model, and the learning and performance assumptions that underlie knowledge tracing are described. The assumptions that underlie knowledge tracing also yield performance predictions. These predictions provide a good fit to students' performance in completing tutor exercises, but a more important issue is how well the model predicts students' performance outside the tutor enviroimient. A previous study showed that the model provides a good fit to average posttest performance across students, but is less sensitive to individual differences. This paper describes a method of individualizing learning and performance estimates on-line in the tutor and assesses the validity of the resulting performance predictions.

The argumentative discourse of computer programmers engaged in a collaborative programming task were analyzed as instances of ecologically valid reasoning behavior. Teams of expert programmers were brought into a laboratory setting to work cooperatively on a software maintenance task. Arguments which occurred spontaneously in the course of the task were examined with respect to: (a) their effect on task performance; and (b) to reveal the sorts of inferential machinery programmers use when they reason with one another. Arguments were found to be important in the formulation of plans as well as the negotiation of strategic priorities with respect to the task. Pragmatic features of the programmers' discourse revealed extensive use of framing devices whose efficacy depended upon interpretation in the context of linked pragmatic scales.

A reasoner may fail at a cognitive task, not because it does not have appropriate knowledge with which to reason, but instead because it does not have the proper index or cue with which to retrieve such knowledge from memory. The reasoner knows this memory item; it simply cannot remember the item. This paper argues that forgetting provides an opportunity for learning through memory reorganization. A reasoner that takes full advantage of such opportunities, however, must be able to reason about its own memory system. To do so, it must possess a language for declaratively representing its reasoning failures and must reflectively inspect such representations if it is to fully explain the reason for its failure. Once such an error is understood as a memory failure, the problem of forgetting is to re-adjust the indexes so that the knowledge is properly retrieved in similar, future situations.

Hyperproof is a computer program created by Barwise and Etchemendy for teaching logic using multimodal graphical and sentential methods, inspired by their theories of heterogeneous reasoning (Barwise and Etchemendy 1994). Elsewhere, we have proposed a theory of the cognitive impact of assigning information to different modalities (Stenning and Oberlander 1992). Our view is that where diagrams are advantageous, it is because they enforce the representation of information, leading to weak expressiveness, thereby facilitating inference. The present study tests and develops these claims by comparing the effects of teaching undergraduate logic classes using Hyperproof and a conUx)l syntactic teaching method. Results indicate that there is significant transfer from the logic courses to logical and analytical reasoning problems. There are also significant interactions between theoretically motivated pre-course aptitude measures and teaching mediod; the interactions influence postcourse reasoning performance in transferdomains. Hyperproof boosts students previously weak on items which benefit from diagram use, whereas the syntactic course appears to degrade the same group of students' graphical strategies. As well as being theoretically interesting, these results provide support for the important practical conclusion that individual differences in aptitude should be taken into account in choosing teaching technique.

In recognition paradigms, increasing the number of occurrences or presentation time in a study list of some words improves performance on these words (the item strength effect), but does not affect the performance on other words {null list strength effect). In contrast, adding new items results in a deterioration of performance on the other words {list length effect). Taken together these results place strong constraints on models of recognition memory. To explain these data an account based on optimisation to the environment is presented. A summary is given of environmental analyses which suggest that (1) the likelihood of recurrence of a word within a context increases as the number of occurrences increases; (2) the repetition rates of other words in a context has no significant effect on the recurrence probability of a word; and (3) the recurrence probability of a word drops as a function of the number of words since the last occurrence of that word. A training set which reflected these constraints was constructed and presented to an optimising connectionist network which was designed to extract recurrence statistics (the Hebbian Recurrent Network). The resultant model is able to model all three of the effects outlined above.

We compared cognitive processing of two complex arithmetic word problems by college students randomly assigned to four different situating tool and social contexts: individualized problem solving with pen and paper; pair problem solving with pen and paper; individualized problem solving on TAPS, a computer-based problem solving tool; and collaborative problem solving on TAPS. Although they solved identical word problems, TAPS users differed from users of conventional tools in that they required relatively more time for problem solving, spent more time in planning activity, and proportionately less time reading. With respect to the influences of social (versus individual) problem solving, collaboration also produced significantly more planning behavior, such that the combined use of TAPS and collaboration produced a marked increase in planning. Also, significantly more behavior associated with metacognitive monitoring occurred in the protocols for pairs. There was no evidence that use of the TAPS tool changed the social nature of the collaboration. However, a qualitative analysis yielded interesting information regarding negotiation processes underlying pair problem solving. For example, we saw specifically some reasons why untrained pair problem solving does not proceed naturally and smoothly. Results are interpreted in terms of situated cognition theory, although symbolic processing theories also can explain much of the data.

A comprehension-based approach to learning assumes that incoming information and background knowledge are integrated to form a mental representation which is subsequently used to incorporate new knowledge. We demonstrate that this approach can indicate when people will learn from instructions. Specifically, we show that a computational model based on the construction-integration theory of comprehension (Kintsch, 1988) can explain and predict how individual users will comprehend help prompts that guide their generation of successful complex commands within an operating system.

In this paper we report preliminary results on how people revise or update a previously held set of beliefs. When intelligent agents learn new things which conflict with their current belief set, they must revise their belief set. When the new information does not conflict, they merely must update their belief set. Various AI theories have been proposed to achieve these processes. There are two general dimensions along which these theories differ: whether they are syntactic-based or model-based, and what constitutes a minimal change of beliefs. This study investigates how people update and revise semantically equivalent but syntactically distinct belief sets, both in symbolic-logic problems and in quasi-real-world problems. Results indicate that syntactic form affects belief revision choices. In addition, for the symbolic problems, subjects update and revise semantically-equivalent belief sets identically, whereas for the quasi-real-world problems they both update and revise differently. Further, contrary to earlier studies, subjects are sometimes reluctant to accept that a sentence changes from false to true, but they are willing to accept that it would change from true to false.

This paper reports research into how mathematical explanations are constructed during conversation based on videotapes of pairs of student math teachers collaboratively writing explanations in geometry. In particular, we analyzed how disagreements about parts of their explanations were managed in these conversations. In contrast to research on disagreement in everyday conversation, explanation disagreements were more likely to overlap with preceding turns and to be stated baldly without prefaces, token agreements or qualifications. However, the observed frequencies of different kinds of disagreements were not consistent with a model favoring explicit substantive disgreement either. Instead, it is proposed that both the interpersonal concerns that would motivate a preference for agreement and the conceptual concerns for a quality explanation that would motivate a preference for substantive disagreement are being managed by participants. Disagreements are co-constructed, and conversants are seen to jointly employ complex devices for introducing and managing disagreement across turns that can satisfy both kinds of concerns with much less conflict betweeen them than might have been expected.

A unique apparatus recorded eye and head movements of subjects as they tapped or only looked at sequences of 2, 4 or 6 nearby, 3-D targets. Each sequence was repeated 10 times to allow an opportunity for learning. A stereotypical pattern of movements was established after 2-3 repetitions. Subjects almost always looked at each target just before tapping it. Looking-only was more difficult than tapping in that it took more time and, unlike tapping, usually did not benefit from practice. The number of targets in a sequence affected timeA^get in both tasks. Sequence length and practice effects show that memory was involved. The persistent strategy of looking before tapping and the subjects' inability to tap a well-leamed patten with eyes closed, show that visual cues were also important We conclude that motor plarming occurred first at the level of the task and then at the level of specific motor programs. The relative difficulty of the less natural, looking-only task, in which the eyes worked without a meaningful cognitive or motor purpose, suggests that efficient eye movement programming requires a natural task of the kind eye movements evolved to serve.

Students often rely on prior work or previously studied examples to help them solve their current problems. In this paper we investigate the relative contributions of easily accessed superficial similarity and deep, solution relevant, structural similarity to memory for prior problems. Some models of memory for analogy suggest that superficial similarity initially selects or constrains memory for prior examples and predicts that analogs that share both surface and structural similarities will be more likely noticed by novices. An experiment is reported in which subjects are observed as they learn how to program. We find that people remember the examples that are related in terms of structural features alone as frequently as those that are related in terms of both structural and superficial features but there is no advantage to having superficial similarities as well. Moreover, even though superficial features sometimes are associated with helpful similarities and sometimes associated with unhelpful similarities people still do not get misled by superficial similarity when that is the only basis for similarity. This finding suggests that models that require superficial similarity as a major selection procedure for analogical reminding may need to be modified for conditions in which people are learning a new skill.

Analogy has always been considered a mechanism for interrelating distinct parts of the world, but it is perhaps just as important to consider how analogy might be used to break the world into comprehensible parts. The MAGI program uses the Structure-Mapping Engine (SME) to flexibly and reliably match a description against itself. The resulting mapping pulls out the two maximally consistent parts of the given description. MAGI then divides out the parts of the mapping and categorizes the mapping as symmetrical or regular. These parts may then be used as the basis for new comparisons. W e theorize that MAG I models how people use symmetry and regularity to facilitate the encoding task. W e demonstrate this with three sets of examples. First, we show how MAGI can augment traditional axis detection and reference frame adjustment in geometric figures. Next, we demonstrate how MAGI detects visual and functional symmetry in logic circuits, where symmetry of form aids encoding synunetry of hmction. Finally, to emphasize that regularity and symmetry detection is not simply visual, we show how MAG I models some aspects of expectation generation in story imderstanding. In general, MAGI shows symmetry and regularity to be not only pretty, but also cognitively valuable.

Contextual and pragmatic knowledge facilitates the eventual interpretation of a syntactically ambiguous sentence. However, psycholinguistic studies have not provided a clear answer to when and how this non-syntactic knowledge is used. One explanation for the discrepancy of the results is that the predictions for parsing processes in context carmot be specified unless they are based on a theory of text comprehension. The constructionintegration model of discourse comprehension (Kintsch, 1988) is proposed as an example for such a theory. The model is parallel and weakly interactive, and its psychological validity has been shown in a variety of applications. Three simulations for syntactic ambiguity resolutions are presented. In the first, syntactic constraints are used to account for the correct interpretation of a garden-path sentence, as well as for common misparses. In the second example, pragmatic knowledge is used to disambiguate a prepositional phrase attachment. In the final example, it is shown that the model can also account for effects of discourse context in the resolution of prepositional phrase attachment ambiguities.

Two experiments are reported to test whether the location of prepositional phrase attachment can be influenced by syntactic and contextual factors. The first experiment tested the hypothesis that attachment is delayed until the word after the prepositional phrase. Replicating the results of Taraban and McClelland (1988), this experiment showed that sentence bias rather than syntactic structure determines the ease of processing; attachment effects were observed on the words after the noun filler. In addition, using sentences in which the noun filler consisted of a compound noun, we also found evidence for delayed attachment. Using sentences in which the noun filler was modified by an adjective, we found evidence for early attachment. In the second experiment, we used context paragraphs to induce earlier attachment for the compound noun sentences. When the first noun of the compound was mentioned in the prior discourse, attachment effects were observed on the disambiguating noun filler. When the first noun was not mentioned, attachment effects were observed, as in Experiment 1, on the words after the prepositional phrase. Thus, the study supports the idea of a contextdependent delay strategy for prepositional phrase attachment.

There have been various suggestions about how children might acquire a proto-ciassification of elements of natural language, such as is conjectured to be necessary to allow the child to "bootstrap" language acquisition (Maratsos 1979; Pinker 1984). One, proposed by Kiss (1972) and Maratsos (1979), but criticised by Pinker (1984), is that children look for distributional correlations between simple linguistic phenomena in the lemguage they heax in order to derive more sophisticated abstract linguistic classifications. Finch & Chater (1992) showed that a relatively complete syntactic classification of the lexicon could be found for common words in natured language using distributional bootstrapping. This paper reviews some of the cirguments Pinker raises against distributional methods, and then describes a system which overcomes his objections, where sequences of words are classified into phrasal classes by a linguisticadly naive statistical aneilysis of distributional regularities fi-om a large, noisy, untagged corpus. For many classes, such as sentence and verb phrase, the accuracy of the classification (ie. the proportion of putative sentences which can in fact be linguistically interpreted as sentences) is in the region of 90%, thus enabling the child to break the "bootstrapping problem".

Identification performance was measured for letters which were briefly presented at different spatial locations and time delays relative to the beginning of manual movement preparation. Identification performance depended on the complexity of the upcoming movement and decreased prior to movement onset. Further findings of similar identification performance with different spatial relations between probe location and manual movement direction cast doubt on the generality of a premotor theory of attention.

Many cognitive tasks involving analogy, such as understanding metaphors, problem-solving, and learning, require the ability to extend mappings as new information is found. This paper describes a new version of SME, called I-SME, that operates incrementally. I-SME is inspired by Keane's lAM model and the use of incremental mapping in Falkenhainer's PHINEAS learning system. W e describe the I-SME algorithm and discuss tradeoffs introduced by incremental mapping, including parallel versus serial pnxessing and pragmatic influences. The utility of 1-SME is illustrated by two examples. First, we show that I-SME can account for the psychological results found by Keane on a serial version of the Holyoak & Thagard attribute mapping task. Second, we describe how I-SME is used in the Minimal Analogical Reasoning System (MARS), which uses analogy to solve engineering thermodynamics problems.

Connectionist accounts of inflectional morphology have focussed on the domain of the English Past Tense (e.g. Rumelhart & McClelland 1986; Plunkett & Marchman 1993). In this inflectional domain, the default mapping process (add /ed/) reflecLs the process of suffixation adopted by the majority of the forms in the language. Connectionist models exploit the imbalance between EngHsh regular and irregular verbs when learning the past tense and when responding to novel forms in a default fashion. Not all inflectional systems have a default mapping which is characterized by a majority of forms in the language. The Arabic Plural System has been cited (Marcus et al. 1993) as one such system where a minority default mapping process operates. The Sound Plural in Arabic applies to only a minority of forms in the lexicon (~104;), yet it appears to adopt the role of a default mapping for novel nouns. W e describe a connectionist model that can learn a minority default mapping analogous to the Arabic plural and discuss its performance in relation to type and token frequency effects, and their distribution within phonetic space.

How does the visual system generate percepts of moving forms? H ow does this happen when the forms are emergent percepts (such as illusory contours or segregated textures) and the motion percept is apparent motion between the emergent forms? A neural model of form-motion interactions is developed to explain parametric properties of psychophysical motion data and to make predictions about the parallel cortical processing streams VI -> M X and VI ->^ V 2 -> MT. The model simulates many parametric psychophysical data arising from form-motion interactions. A key linkage between form and motion data is articulated in terms of properties of visual persistence and properties of apparent motion. The model explains how an illusory contour can move in apparent motion to another illusory contour or to a luminance-derived contour; how illusory contour persistence relates to the upper ISl threshold for apparent motion; and how upper and lower ISI thresholds for seeing apparent motion between two flashes decrease with stimulus duration and narrow with spatial separation (Korte's laws). Psychophysical data are derived from an analysis of how orientationally tuned form perception mechanisms and directionally tuned motion perception mechanisms interact to generate consistent percepts of moving forms.

It is well known that when a connectionist network is trained on one set of patterns and then attempts to add new patterns to its repertoire, catastrophic interference may result. The use of sparse, orthogonal hidden-layer representations has been shown to reduce catastrophic interference. The author demonstrates that the use of sparse representations may, in certain cases, actually result in worse performance on catastrophic interference. This paper argues for the necessity of maintaining hidden-layer representations that are both as highly distributed and as highly orthogonal as possible. The author presents a learning algorithm, called context-biasing, that dynamically solves the problem of constraining hiddenlayer representations to simultaneously produce good orthogonality and distributedness. On the data tested for this study, context-biasing is shown to reduce catastrophic interference by more than 50% compared to standard backpropagation.

Cross-modal priming experiments have shown that surface variations in speech are perceptually tolerated as long as they occur in phonologically viable contexts. For example, [klim] {cleam) gains access to the mental representation of clean when in the context of [klimpoks] (cleam parks), since the change is a natural one, reflecting the phonological process of place assimilation. This implies that speech perception involves processes of phonological inference, which recover the underlying form of speech. Here we investigate the locus of these inference processes, using the phoneme monitoring task. A set of stimulus sentences were created containing deviations that were either phonologically viable (as in cleain parks above) or unviable. In Experiment 1, subjects monitored for the segment underlying the surface change (in the above example, /n/) and in Experiment 2 the following segment (/p/) was the taiget. In addition, the lexical status of the carrier word was manipulated (e.g., clean vs threan), contrasting lexical and non-lexical theories of phonological inference. Both experiments showed stiong effects of phonological viability for real words, with weaker effects for the non-word stimuli. These results suggest that phonological inference can occur non-lexically, but that it interacts strongly with the process of lexical access.

Three experiments are reported that examine the impact of people's goals and conceptual understanding on graph interpretation, in order to determine how people use graphical representations to evaluate functional dependencies between continuous variables. Subjects made inferences about the relative rate of two continuous linear variables (altitude and temperature). We varied the assignments of variables to axes, the perceived cause effect relation between the variables, and the causal status of the variable being queried. The most striking finding was that accuracy was greater when the Slope-Mapping Constraint was honored, which requires that the variable being queried - usually the effect or dependent variable, but potentially the cause instead — is assigned to the vertical axis, so that steeper lines map to faster changes in the queried variable. This constraint dominates when it conflicts with others, such as preserving the low-level mapping of altitude onto the vertical axis. Our findings emphasize the basic conclusion that graphs are not pictures, but rather symbolic systems for representing higher-order relations. We propose that graphs provide external instantiations of intermediate mental representations, which enable people to move from pictorial representations to abstractions through the use of natural mappings between perceptual properties and conceptual relations.

Directional asymmetiy is a well-documented phenomenon in research on similarity, meuphor, and analogy. In this paper, we present an account of this phenomenon based on structural alignment. We propose that a major source of asymmetry is coherence imbalance: that is, a difference in the degree of systematicity of the relational structures being compared. These claims are tested in three experiments which examine the relationship between asymmetry, informativity, and conceptual coherence. The results support the hypodiesis that coherence imbalance is a key factor in directional comparison processes. Further, by incorporating the insights offered by structural alignment, coherence imbalance advances a more functional account of asymmetry.

The recent availability of large on-line parsed corpora makes it possible to test theories of psycholinguistic complexity by comparing the frequency distributions of closely related constructions. In this paper, we use this technique to test the psycholinguistic theory proposed by Gibson et al. (1993), which includes two independently motivated attachment principles: Recency Preference and Predicate Proximity. In order to test this theory, we examined two general classes of attachment ambiguities from the parsed Wall Street Journal corpus from the Penn Treebank: 1) ambiguities which involve three prospective noun phrase attachment sites; and 2) ambiguities which involve three prospective verb phrase attachment sites. Given three prospective noun phrase (NP) sites in English, the theory most naturally predicts a complexity ordering of NP3 (easiest, most recent), NPi, NP2, but a ranking of VP3, VP2, VPi for verb phrase attachments. Our corpus analyses support both of these predictions.

Cognitive scientists have often pondered the question of perceptual spaces, that is, the question of how a certain gamut of familiar stimuli might be organized in the mind. W e present Trajectory Mapping as an alternative clustering method to the traditional algorithm of Multi-Dimensional Scaling. We suggest that given data about the relationships among stimuli, Multi-Dimensional Scaling provides the one type of information (geometric), while Trajectory Mapping offers a second type (relational). As an illustration we present the initial results of applying both clustering techniques to subjects' perceptions of musical intervals. While an interpretation of the MultiDimensional Scaling requires a priori knowledge of music theory, Trajectory Mapping directly reveals the music theory that has been internalized by subjects.

Do children acquire rules for main stress assignment or do they learn stress in an exemplar-based way? In the language acquisition literature, the former approach has been advocated without exception: although they hear most words produced with their appropriate stress pattern, children are taken to extract rules and do not store stress patterns lexically. The evidence for a rule-based approach is investigated and it will be argued that in the literature this approach is preferred due to an extremely simplified interpretation of exemplar-based models. W e will report experiments showing that Instance-Based Learning, an exemplar-based model, makes the same kinds of stress related errors in production that children make: (i) the amount of production errors is related to metrical markedness, and (ii) stress shifts and errors with respect to the segmental and syllabic structure of words typically take the form of a regularization of stress patterns. InstanceBased Learning belongs to a class of Lazy Learning algorithms. In these algorithms, no explicit abstractions in the form of decision trees or rules are derived; abstraction is driven by similarity during performance. Our results indicate that at least for this domain, this kind of lazy learning is a valid alternative to rule-based learning. Moreover the results plead for a reanalysis of language acquisition data in terms of exemplar-based models.

To date, the major focus of research in knowledge representations for artificial intelligence has been on sentential or linguistic formalisms involving logic and rulebased reasoning. There is a growing body of evidence suggesting, however, that much of human problem solving is achieved, not through the application of rules of inference, but rather through the manipulation of mental models. Such a model is represented by a system with a similar relational structure to the reality it represents. Moreover, spatial reasoning with models involves the inspection and transformation of representations in ways that are analogous to visually inspecting and physically transforming entities in the world. Since a crucial component of knowledge acquisition is to capture an expert's mental state and reasoning strategies, it is important to shift some of the attention of AI research to the study of representation techniques that correspond to the mental models used by humans. The paper begins with a cognitive perspective on model-based reasoning. A knowledge representation scheme for spatial reasoning with models is then presented. In this scheme, which has evolved from research in computational imagery, spatial models are represented as symbolic arrays where dimensions of the array correspond to transitive order relations among entities.

Much previous work in developing computational models of scientific discovery has concentrated on the formation of basic laws. The important role played by additional assumptions in this process is a neglected research topic. W e argue that hypotheses about structure are an important source of such additional assumptions, and that knowledge of this type can be embodied in the notion of Informal Qualitative Models (IQMs). In this paper, we demonstrate that such models can be synthesised by applying a set of operators to the most fundamental model in a domain. Heuristics are employed to control this process, which forms the basis of an architecture for model-driven scientific discovery. Conventional data-driven discovery techniques can be integrated into this architecture, resulting in laws which depend crucially on the model that is applied to a problem. This approach is illustrated by an historical survey of eighteenth and nineteenth century solution chemistry, which focuses on the evolution of the models employed by scientists. A series of models are synthesised which reflect these historical developments, showing the importance of structural models both in understanding certain aspects of the scientific discovery process, and as a basis for practical discovery systems.

A key problem in cognitive science concerns how the brain binds together parts of an object into a coherent visual object representation. One difficulty that this binding process needs to overcome is that different parts of an object may be processed by the brain at different rates and may thus become desynchronized. Perceptual framing is a mechanism that resynchronizes cortical activities corresponding to the same retinal object A neural network model based on cooperation between oscillators via feedback from a subsequent processing stage is presented that is able to rapidly resynchronize desynchronized featural activities. Model properties help to explain perceptual framing data, including psychophysical data about temporal order judgments. These cooperative model interactions also simulate daU concerning the reduction of threshold contrast as a function of stimulus length. The model hereby provides a unified explanation of temporal order and threshold contrast data as manifestations of a cortical binding process that can rapidly resynchronize image parts which belong together in visual object representations.

This research investigated the effects of prior knowledge on learning in psychologically-plausible connectionist networks. This issue was examined with respect to the benchmark orthography-to-phonology mapping task (Sejnowski & Rosenberg, 1986; Seidenberg & McClelland, 1989). Learning about the correspondences between orthography and phonology is a critical step in learning to read. Children (unlike the networks mentioned above) bring to this task extensive knowledge about the sound-structure of their language. We first describe a simple neural network that acquired some of this phonological knowledge. W e then summarize simulations showing that having this knowledge in place facilitates the acquisition of orthographicphonological correspondences, producing a higher level of asymptotic performance with fewer implausible errors and better nonword generalization. The results suggest that connectionist networks may provide closer approximations to human performance if they incorporate more realistic assumptions about relevant sorts of background knowledge.

This paper describes a lexical acquisition mechanism that was implemented in order to increase the robustness of a Natural Language Processing system. Although the mechanism was not intended to be a cognitive model of children's language acquisition, it demonstrates many similarities with psycholinguistic findings. In particular, the structure of the domain knowledge representation forces the system to take a bipolar approach to learning nouns and verbs. Psycholinguistic studies demonstrate differing treatment of nouns and verbs by children and suggest a structural basis for this difference. The knowledge-level similarities between our system and human linguistic knowledge make it possible to infer that children must adopt a similar strategy to effectively learn word meanings.

This paper presents empirical evidence regarding the nature of our commonsense concept of behef. The findings have significant bearing upon claims made by authors concerned with the Folk Psychology Debate—in particular, they challenge Stephen Stich's (1983) claim that folk psychology is committed to a broad account of belief states. In contrast it is found that folk psychology favours a narrow account of belief. This result is important in refuting Stich's claim that the folk psychological concept of belief has no role to play in a developed cognitive science. The paper also presents evidence regarding the influence of several factors on folk psychological judgements of belief individuation (emphasised similariiies/differences between the referents of beliefs, nature of past beliefs, goal of classification), and introduces a methodology by which to investigate further factors. It is argued that the observed conflict between individual speculations about likely folk psychological intuitions within the philosophical literature and actual empirical data regarding subjects' responses highlights the important contribution of experimental psychology in exploring such philosophical issues.

The psychological validity of inheritance reasoners is clarified. Elio and Pelletier (1993) presented the first pilot experiment exploring some of these issues. W e investigate other foundational assumptions of inheritance reasoning with defaults: transitivity, blocking of transitivity by negative defaults, preemption in terms of strucUirally defined specificity and structurally defined redundancy of information. Responses were in accord with the assumption of at least limited transitivity, however, reasoning with negative information and structurally defined specificity conditions did not support the predictions of the literature. 'Preemptive' links were found to provide additional information leading to indeterminacy, rather than providing completely overriding information as the literature predicts. On the other hand, results support the structural identification of certain links as redundant. Other findings suggest that inheritance proof-theory might be excessively guided by its syntax.

This paper presents a way that enables robots to learn abstract concepts from sensory/perceptual data. In order to overcome the gap between the low-level sensory data cind higher-level concept description, a method called feature abstraction is used. Feature abstraction dynamically defines abstract sensors from primitive sensory devices and makes it possible to learn appropriate sensory-motor constraints. This method has been implemented on a reed mobile robot as a learning system called ACORN-II. ACORN-II was evaluated with some empirical results eind shown that the system can learn some abstract concepts more accurately than other existing systems.

The Theory of Explanatory Coherence, or TEC, (Ranney & Thagard, 1988; Thagard. 1989, 1992) and ECHO, a connectionist implementation of TEC, attempt to model human reasoning about evidence and hypotheses. The ECH O model is based on the simultaneous satisfaction of multiple constraints. This yields predicted activations ("believabilities") for propositions, which are based on the propositions' evidential status, their explanatory relationships, and their contradictory relationships. While ECH O has been demonstrated to usefully model human reasoning, it does not model processing limitations on the maintenance of coherence. WanderECHO is a variation on the ECH O model that attempts to simulate attentional and memorial limitations with a stochastic updating algorithm that is based on a traveling focus of attention. Several variants of the WanderECHO simulation were applied to Schank and Ranney's (1991) data, and were found to generally simulate subjects' mean believability ratings better than standard ECHO.

This paper outlines an implemented system called PROVERB that explains machine-found natural deduction proofs in natural language. Different from earlier works, we pursue a reconstructive approach. Based on the observation that natural deduction proofs are at a too low level of abstraction compared with proofs found in mathematical textbooks, we define first the concept of socalled assertion level inference rules. Derivations justified by these rules can intuitively be understood as the application of a definition or a theorem. Then an algorithm is introduced that abstract machine-found ND proofs using the assertion level inference rules. Abstracted proofs are then verbalized into natural language by a presentation module. The most significant feature of the presentation module is that it combines standard hierarchical text planning and techniques that locally organize argumentative texts based on the derivation relation under the guidance of a focus mechanism. The behavior of the system is demonstrated with the help of a concrete example throughout the paper.

Synchrony of firing has recently become a popular technique for dynamic binding in neural networks, and has been applied to numerous problem domains. However, hierarchical structures are difficult to represent using synchrony for binding. This paper presents our progress toward a framework for representing hierarchies in a neural network using synchrony for dynamic binding. We illustrate the approach with a model of analogical mapping. The model (IMM2) uses synchrony to bind case roles to objects within propositions. Hierarchies are established by allowing units representing propositions to play a dual role, acting both as the argument of one proposition and as a pointer to another.

When a wheel rolls along a flat surface, a point on the wheel's perimeter follows a cycloid trajectory. Subjects, however draw the curtate cycloid, characterized by bottom loops, rather than the cycloid to depict the path that a point on a static wheel's perimeter would trace if the wheel were rolling. This is the curtate cycloid illusion. In Experiment 1, we show that animating the wheel does not dispel the illusion and that subjects high in spatial ability are less susceptible to the illusion than are low-spatials. Experiments 2, 3a, and 3b supported the hypothesis that the illusion occurs when subjects reallocate cognitive resources from processing a rolling wheel's translation to computing its instant centers, the point about which the wheel is rotating at a given instant in time. This reallocation occurs only when a reference point on the wheel's perimeter contacts and leaves the surface. We conclude that the illusion does not reflect fundamental perceptual biases, but rather stems from transient shortages of cognitive resources during the higher-level processing of the wheel's translation and rotation.

This paper presents a computational model for stereopsis. Laplacian of Gaussian filters are used to simulate ganglion cells and LGN cells and zero-crossings extracted provide spatial features in the visual scene. A set of one-octave Gabor filters is used to extract orientation information, which cover 0 to 60 cycles/degree interval in the human visual system. A Gaussian sphere model is used to map a 3D space onto two 2D image planes, which combines monocular cues with binocular cues in stereo matching. The determinant of the Jacobian of the mapping is derived and matching is performed using zerocrossings associated with their orientation information. The possibility of transferring the knowledge such as the probability of occurrence of visual scenes to the matching process from the mapping is discussed. Relaxation labelling is used as a co-operative process, which simulates binocular fusion and rivalry in the human visual process.

Agents in a dynamic world must continue to comprehend and reason about events, even after they learn that previously encoded information about an event is incorrect. As a result, some mechanism is needed to modify incorrect information in memory, and allow one to use new, superceding knowledge instead. Ho w is misinformation suppressed in human memory? A study using a text understanding paradigm and a standard anaphoric inference task investigates this problem of updating memory. Subjects read a set of stories, half of which contained a conection, and were asked to make a speeded wordrecognition judgment for a probe word appearing after an anaphor sentence. Subjects in a short delay condition showed slower reaction times to correct referents in correction stories than in control stories that did not contain misinformation. Those in the longer delay condition showed no difference in reaction times to correct referents, but more priming for invalidated items in correction stories. These results suggest that misinformation can interfere with accessing correct information, but that an additional comprehension process, possibly suppression-like, may facilitate access to correct information after delay.

Abduction is the process of constructing a plausible explanation for a set of observations. It is the fundamental type of reasoning in many complex tasks such as scientific discovery and diagnosis. This paper presents a mental-model theory of human abductive skill and its acquisition in which abduction is viewed as the sequential comprehension and integration of daU into a single situation model. Comprehension and integration are accomplished using satisficing search of multiple problem spaces. The model has been implemented in Soar and has been tested by comparing its predictions to those of human subjects. The experimental results show that the model can account for several important behavioral regularities, including power-law speed-up, bow the order of data presentation affects a response, deviation of responses from probability theory, and bow the task and domain characteristics affect a person's response.

When faced with a novel problem, people can sometimes decide what to do by imagining alternative sequences of actions and then taJdng the sequence that solves the problem. In many problems, however, various constraints, such as working memory capacity, limit the amount of internal lookahead that people can do. This paper describes Bottom-Up Recognition Learning (BURL), a model of limited-lookahead learning based on final first learning and knowledge compilation. In BURL, knowledge compilation of limited-lookahead search over successive problemsolving trials transfers knowledge from the leaf nodes of a problem space to the top node. Two experiments test BURL'S predictions. The first compares the Soar implementation of BURL to human subjects learning to play two Tlc-Tac-Toe isomorphs. This experiment shows that BURL can account for learning that occurs when subjects can perform a limited lookahead. The second experiment studies transfer between two strategy acquisition tasks for one isomorph. This experiment shows that BURL must be used in conjunction with other learning methods to fully explain skill acquisition on limited lookahead tasks.

We describe an algorithm that constructs mental models of assertions containing sentential connectives, such as and. if, and or. It performs at three levels of expertise depending on the completeness of the models it constructs. At a rudimentary level of performance, it constructs models that make explicit as little as possible. One unexpected consequence is that it produces the same explicit models for assertions of the form: if p then q, and if r then s if p then q, 21 if r then s p and q, 01 r and s. W e initially suspected that there was a bug in the algorithm (or theory), but there was not. W e therefore carried out two experiments with logically-untrained subjects. Their results confirmed the phenomena: for many individuals, a conjunction of conditionals is equivalent to their disjunction, which in turn is equivalent to a disjunction of conjunctions.

Two experiments in category learning are used to examine two types of categorization models. In both a two and four choice experiment, subjects are shown to fail to learn to optimally classify two dimensional stimuli. The general recognition theory (CRT) of Ashby & Maddox (1990) predicts quadratic decision bounds. The first experiment disconfirms this. The extended GR T predicts that learners adopt a bound of complexity equivalent to the optimal one. The second experiment disconfirms this as well. Both experiments support the idea that general resources of adaptive systems can provide explanations of observed sub-optimal behavior.

Design is a complex cognitive task that pushes the limits of human information processing. How do expert designers handle this complexity? Professional and student architects solved a real-world diagram construction task that required satisfying multiple, sometimes conflicting, constraints to achieve an acceptable design. Professionals' initial designs were more consistent with task constraints and remained more consistent throughout problem solution. Students restructured their designs more often in their unsuccessful attempts to satisfy the multiple constraints imposed by the task. Analysis of subjects' verbal and action protocols suggests that one aspect of professionals' superior performance is their early recognition of the critical constraints on a design. Professionals handle these constraints before others to structure the remaining, more negotiable, constraints. By properly ordering constraints, professionals effectively minimize constraint conflicts. As conflict resolution has high processing costs, constraint prioritization may be one way that professionals cope with the complexity of design.

In any given analogy, there are potentially a large number of possible mapping interpretations. One of the key issues in analogy research is how one of these mappings comes to be selected as optimal and used as the basis for the analogical comparison. It is well-established that structural factors, notably systematicity, can act as selection constraints on mapping. The present work tests to see if pragmatic and adaptation factors can also act as selection constraints on mapping. The selection of a mapping based on pragmatic factors proposes that people can exploit the higher-order, schematic structure of a domain to select one mapping over another. With respect to adaptation factors, the proposal is that a mapping will be selected if it is evaluated as being easily adapted relative to other competing mappings. Both of these predictions are tested in a novel, problem solving paradigm. The main finding is that adaptation factors do act as a selection constraint but that pragmatic factors do not. The implications of these results for computational models of analogy are discussed.

Two experiments assessed the membership functions that Germ an speakers assign to 12 adverb phrases and 17 modal verb fonns that express probability assessments. These expressions fall largely into three rather homogeneous classes. The membership functions are used as part of the semantic knowledge base of the natural language dialog system Pragma, one of whose purposes is to model pragmatic and contextual influences on the use of vague expressions. The system's normative model accounts for the role, in the selection and interpretation of vague probability expressions, of the listener's prior expectations, the speaker's dialog motivation, and the expressions that the speaker could have used but did not.

Context-sensitivity is an important characteristic feature of every cognitive process and therefore should be reflected in every architecture pretending to explain human cognition. In this paper some experimental facts demonstrating context effects on various cognitive processes are reviewed and an attempt at context modeling is described A hybrid (symbolic/connectionist) cognitive architecture, DUAL , is proposed. It consists of a multitude of agents having both a symbolic and a connectionist part. The symbolic part represents some knowledge structure, while the coimectionist part represents its relevance to the current context The performance of the cognitive system emerges as result of the work and interaction of the currently active agents, where the set of active agents is not predefined for a specific task but is dynamic and reflects the specific context. So particular symbolic operations and data structures may be supported or suppressed depending on the particular activation pattern of the connectionist parts which represent the context-dependent relevance of the operations and structures. In this way a context-sensitive computation emerges. A n example of context-sensitive deductive reasoning is described.

Cluster analysis has been successfully applied to the problem understanding hidden unit representations in both feed-forward and recurrent neural networks. While the topological properties of feed-forward networks may support the use of cluster analysis, the results described within this paper suggest that applications to recurrent networks are not justified. This paper illustrates how clustering fails to provide useful insights into the underlying task-dependent information processing mechanism of recurrent networks. In this paper, I first demonstrate that randomly generated networks display a surprising amount of clustering before training. Then I explain that the clustering structure emerges, not in response to the task training, but because of the volume-reducing iterated mappings that comprise the commonly used recurrent neural networks models.

Theorists of human learning, in domains as various as category learning and language acquisition, have grappled with the issue of whether learners induce rules or remember exemplars, or both. In this article we present new dau that reflect both rule induction and exemplar encoding, and we present a new connectionist model that specifies one way in which rule-based and exemplar-based mechanisms might interact Our empirical study was motivated by analogy to past tense acquisition, and specifically by the previous work of Palermo and Howe (1970). Human subjects learned to categorize items, most of which could be classified by a simple rule, except for a few frequently recurring exceptions. The modeling was motivated by the idea of combining an exemplar-based module (ALCOVE, Kruschke, 1992) and a rule-based module in a connectionist architecture, and allowing the system to learn which module should be responsible for which instances, using the competitive gating mechanism introduced by Jacobs, Jordan, Nowlan, and Hinton (1991). We report quantitative fits of the model to the learning data.

This work examines the contribution of shape features to subjects' judgments of typicality for visual categories. Shape was found to make a strong contribution to typicality, as evidenced by the strong correlation between results on pictures and those on silhouettes of the same pictures. Also, di^erent measures of the contribution of shape template overlap, compactness, and number of parts - were shown to capture different aspects of that contribution. As one of the fundamental problems in category research is to determine the features used in categorization (e.g., Medin, 1989), the current work is important because it makes progress on this problem.

A recurrent network was trained from sentence examples to construct symbolic parses of sentence forms. Hundreds of sentences, representing significant syntactic complexity, were formulated and then divided into training and testing sets to evaluate the ability of a recurrent network to learn their structure. The network is shown to generalize well over test sentences and the errors that do remain are found to be of a single type and related to human limitations of sentence processing.

A common method of teaching vocabulary involves presenting students with new words in context and having the students derive the meaning of these words based on contextual cues. Beck, McKeown and McCaslin (1983) have argued that the contexts used to teach new words should be highly constraining. Although highly constraining contexts avoid ambiguity they do not present the learner with the necessity of combining contextual and word specific information and thus practicing skills needed for general comprehension. W e suggest that a superior method of teaching is to relax the amount of contextual constraint because to optimize the learning from the presentation of a sentence the student must use both top down and bottom up processes to discover the meaning of the sentence, thus integrating two sources of knowledge about the word. The present research compares knowledge and use of newly learned words between students who learned the new words using three encounters with highly constraining contexts, three encounters with moderately constraining contexts or three progressively less constraining contexts. Students were given definitional and comprehension tests both immediately after study and at a one week delay. The results suggest that repeated encounters with moderately constraining contexts are superior to repeated encounters with highly constraining contexts.

Recent models of metrical structure in music rely upon notions of oscillation and synchronization. Such resonance models U-eat the perception of metrical structure as a dynamic process in which the temporal organization of musical events synchronizes, or entrains, a listener's internal processing mechanisms. The entrainment of a network of oscillators to an afferent rhythmic pattern models the perception of metrical structure. In this paper, 1 compare one resonance model with several previously proposed models of meter perception. Although the resonance model is consistent with previous models in a number of ways, mathematical analysis reveals properties that are quite distinct from properties of the previously proposed models.

Current theories and supporting simulations of similaritybased retrieval disagree in their process model of semantic similarity decisions. We compare two current computational simulations of similarity-based retrieval, MAC/FA C and ARCS, with particular attention to the semantic similarity models used in each. Four experiments are presented comparing the performance of these simulations on a common set of representations. The results suggest that MAC/FAC, with its identicality-based ccmstraint on semantic similarity, provides a better account of retrieval than ARCS, with its similarity-table based model.

Much recent research on modeling memory processes has focused on identifying useful indices and retrieval strategies to support particular memory tasks. Another important question concerning memory processes, however, is how retrieval criteria are learned. This paper examines the issues involved in modeling the learning of memory search strategies. It discusses the general requirements for appropriate strategy learning and presents a model of memory search strategy learning applied to the problem of retrieving relevant information for adapting cases in case-based reasoning. It discusses an implementation of that model, and, based on the lessons learned from that implementation, points towards issues and directions in refining the model.

ISOR is a neural network system for object recognition and scene analysis that leeims visual schemas from examples. Processing in VISOR is based on cooperation, competition, and parallel bottom-up and top-down activation of schema representations. Similar principles appear to imderlie much of human visual processing, and VISOR can therefore be used to model vairious perceptucd phenomena. This paper focuses on anedyzing three phenomena through simulation with VISOR: (1) priming and mental imagery, (2) perceptual reversal, and (3) circular reaction. The results illustrate similarity and subtle differences between the mechanisms mediating priming and mental imagery, show how the two opposing accounts of perceptual reversal (neural satiation and cognitive factors) may both contribute to the phenomenon, and demonstrate how intentional actions can be graduaJly learned from reflex cictions. Successful simulation of such effects suggests that similar mechanisms may govern human visual perception eind learning of visual schemas.

Over the past twenty years numerous studies have investigated the extent to which morphological constituents of words are activated during the process of word recognition. In the vast majority of these studies it has been assumed that a correspondence exists between the formal linguistic analysis of a word and its representation in the minds of native speakers. This paper investigates the the extent to which this correspondence can be affected by individual variation that is associated with education, exposure and training. We investigated student who had recently completed a course in medical terminology. These students, and matched control subjects, responded to medical and nonmedical multimorphemic stimuli in a lexical decision task. The results indicate that the medical terminology students' training affected their performance on novel medical words as well as their performance on very common medical words (e.g., psychiatry) that would have been part of their vocabulary prior to taking the course. The results therefore support the view that automatic unconscious lexical processing can indeed be modified by explicit training and specialized exposure. This finding has consequences for the generalizability of studies conducted on university students to the general population of native speakers.

Learning the past tense of English verbs has become a landmark task for testing the adequacy of cognitive modeling. W e review a set of intriguing psychological phenomena that any modeling of past-tense acquisition has to account for. Traditional grammatical theories fail to explain phenomena of irregular verbs, while connectionist modek, which require no symbols and explicit rules, fail on regular verbs. W e present a generalpurpose symbolic pattern associator (SPA for short) which learns a set of sufficient and necessary symbolic rules for both distinguishing and predicting regular and irregular verbs. Our aU-rule theory is similar in spirit to Pinker's (1991, 1993) modular hypothesis, and is able to account for most psychological phenomena in past-tense acquisition. Even on the task of irregru/ar past-tense generalization, the SPA is judged to be slightly more plausible than the connectionist model by adult native English speakers. Our results support the view that language acquisition and processing should be better modeled by symbolic, rather than connectionist, systems.

Meeting scheduling takes place when a group of people intend to meet with each other. Since each person has individual availability constraints and preferences, meeting scheduling is naturally distributed and there is a need to schedule the meeting in such a way as to consider the preferences of the set of meeting participants. In addition, individual meeting constraints and preferences may change both as a result of an agent's situation or as a result of other agents' scheduling decisions. Therefore, there is a need for distributed reactive schedule revision in response to changing requirements and constraints. W e present an approach to distributed meeting scheduling based on modeling and communication of constraints and preferences among the agents. When a feasible global schedule cannot be found, agents enter a negotiation and relax their constraints. The approach enables the agents to find and reach agreement on the schedule with the highest joint utility and to reactively revise the schedule in response to new information.

Psychological investigations have led to considerable insight into the working of the human language comprehension system. In this article, we look at a set of principles derived from psychological findings to argue for a particular organization of linguistic knowledge along with a particular processing strategy and present a computational model of sentence processing based on those principles. Many studies have shown that human sentence comprehension is an incremental and interactive process in which semantic and other higher-level information interacts with syntactic information to make informed commitments as early as possible at a local ambiguity. Early commitments may be made by using top-down guidance from knowledge of different types, each of which must be applicable independently of others. Further evidence from studies of error recovery and delayed decisions points toward an arbitration mechanism for combining syntactic and semantic information in resolving ambiguities. In order to account for all of the above, we propose that all types of linguistic knowledge must be represented in a common form but must be separable so that they can be applied independently of each other and integrated at processing time by the arbitrator. W e present such a uniform representation and a computational model called COMPERE based on the representation and the processing strategy.

We describe the perceptual foundations of a sensorimotor model of early childhood phonetic and articulatory development. The model's auditory perception is sensitive to prosodic and syllabic structure and simulates the categorical phonetic perception of late infancy. Importantly, the model relies on exclusively acoustic cues and their statistical distribution in the linguistic environment, avoiding prior assumptions of articulatory-acoustic correlations or linguistic contrasts which are inappropriate for a model of perceptual development. The model detects and categorizes speech segments, which, despite their acoustic basis, correlate with linguistic events and articulatory gestures. The resulting representation supports not only word recognition but also the unique demands of articulatory motor conU-ol and its development. In simulations examining the distinctiveness and faithfulness of the representation, we find that it preserves and makes explicit information about the phonetic properties of the acoustic signal.

A model for lexical disambiguation is presented that is bfised on combining the frequencies of past contexts of ambiguous words. The frequencies axe encoded in the word representations and define the words' semantics. A Simple Recurrent Network (SRN) parser combines the context frequencies one word at a time, edways producing the most likely interpretation of the current sentence at its output. This disambiguation process is most striking when the interpretation involves semantic flipping, that is, an cilternation between two opposing meanings as more words are read in. The sense of throsing a ball alternates between deince zind baseball as indicators such as the agent, location, and recipient are input. The SR N parser demonstrates how the context frequencies are dynamically combined to determine the interpretation of such sentences. We hypothesize that several other aspects of ambiguity resolution are based on similar mechanisms, and can be naturally approached from the distributed connectionist viewpoint.

In this paper, a dynamic niodel of human time perception is presented which treats time as phase, relative to the period of an oscillator that adapts its oscillation rate in response to an input rhythm. The adaptive oscillator mechanism is characterized by four fundamental properties: (1) a preferred oscillation rate which captures the notion of a preferred tempo, (2) a fast-acting synchronisation procedure which models our ability to perceptually lock onto salient aspects of a rhythm, (3) a decay process to oppose synchroniaation, and (4) a drift process which causes the preferred rate to gradually drift towards the adapted rate, thereby modeling the context effects of long-term pattern exposure. By assuming that sensitivity to duration is a function of oscillator entrainment to the contextual rhythm, the model provides a qualitative match to data on tempo discrimination, and predicts the types of errors subjects would make on such tasks. These predictions are in agreement with data showing that subjects overestimate short intervals and underestimate long intervals.

We present the results of a simple experiment in lowercase letter recognition. Unlike most psychology studies of letter recognition, we include in our data set letters at the extremes of their categories and investigate the recognition of letters of multiple typefaces. W e are interested in the relationship between the recognition of normal letters and the recognition of non-standard letters. Results provide empirical evidence for top-down conceptual constraints on letter perception in the form of roles and relations between perceptually-based structural subcomponents. A process model based on the hypothesis developed below is currently being implemented.

The sound-form generation of a word in speech production involves the retrieval of segmental and suprasegmental information from the mental lexicon. A translation task experiment showed that the naming latencies of target items can be reduced when prime words are presented that have the same placement of the lexical stress as the target. However, this reduction will only occur when primes and targets have the same word onset. A second experiment showed that primes that have the same number of segments as the targets will cause naming facilitation compared to primes that have different numbers of segments. I have developed a new model of phonological encoding that incoqwrates ordered selection of the various elements. Lexical stress is chosen fu-st, followed by information about the number of slots, the word onset, the second segment, and the other segments, until all segments have been selected. The model further employs mechanisms that allow for the retrieval of the initial segment to influence the retrieval of lexical stress. Various simulations show that the model can replicate the findings of the two experiments. Other models of phonological encoding largely neglect suprasegmental retrieval and cannot explain these results.

This paper analyzes how mathematicians prove theorems. The analysis is based upon several empirical sources such as reports of mathematicians and mathematical proofs by analogy In order to combine the strength of Uaditional automated theorem provers with human-like capabilities, the questions arise: Which problem solving strategies are appropriate? Which representations have to be employed? As a result of our analysis, the following reasoning su^tegies are recognized: proof planning with partially instantiated methods, structuring of proofs, the transfer of subproofs and of reformulated subproofs. W e discuss the representation of a component of these reasoning suategies, as well as its properties. W e find some mechanisms needed for theorem proving by analogy, that are not provided by previous approaches to analogy. This leads us to a computational representation of new components and procedures for automated theorem proving systems.

Novices often experience great difficulty leeirning new domains. Thus, understanding how best to scaffold novice problem solving has potentiaUy tremendous importance for learning in formal domains. In this paper, we present results from ein experimental study that compared learning outcomes of students solving introductory programming problems in three different learning environments. This range of environments v£iries in two ways. First, the notations used in the environments vary between diagrammatic and textuaJ. More importantly, the environments differ in the cognitive activities students are led to perform while solving problems, such as prediction of intermediate results and noting future goals to achieve. This experiment demonstrated that environments that scaffold more of the important cognitive activities lead to superior performance, regardless of whether the environments are textual or diagrammatic.

This paper addresses category typicality in the context of a category naming task. In contrast to the predominant effort with gradient models, a symbolic search framework is taken. Within this framework, the SC A (Symbolic Concept Acquisition) model demonstrates varying response times as a function of an instance's intra-category typicality. Here its coverage is expanded to inter-category typicality. A functionally motivated extension for SC A is advanced that pursues search backtracking under ambiguous cases. I explain how the backtracking extension accounts for inter-category typicality effects, and support it with some empirical evidence. I discuss how the effect generalizes to a larger class of symbolic search models.

Three studies investigated the role of perceptual and quantitative situational factors on the structure of 5th- and 6th-graders' mental models. A task involved a carton of orange juice made from concentrate and water, and two glasses of different sizes filled from the carton. The children had to predict whether the two glasses would taste the same. W e manipulated whether students were presented with physical, diagrammatic, photographic, or textual information. W e also manipulated the type of relationship specified between quantities: qualitative, easy numerical, or difficult numerical. W e found that for the diagram condition, difficult numerical relationships yielded poor performance, whereas the easy numerical and qualitative relationships yielded excellent performance. In contrast, in the physical condition, the easy numerical relationships yielded poor performance, whereas the difficult numerical and quaUtative relationships yieldedexcellent performance. These and otherresults are interpreted by developing a sketch of the mental models preproportional children construct to reason about this quantitative situation, and describing how situational factors influence the construction of the models. For example, physical features led to models that captured the identity relationship between the juice in the glasses (e.g., the juice came from the same carton) whereas numerical features led to models that captured the relationship between the constituents of concentrate and water in each glass (e.g., within a glass there is more water than concentrate).

Reading has been studied for decades by a variety of cognitive disciplines, yet no theories exist which sufficiently describe and explain how people accomplish the complete task of reading real-world texts. In particular, a type of knowledge intensive reading known as creative reading has been largely ignored by the past research. W e argue that creative reading is an aspect of practically all reading experiences; as a result, any theory which overlooks this will be insufficient. W e have built on results from psychology, artificial intelligence, and education in order to produce a functional theory of the complete reading process. The overall framework describes the set of tasks necessary for reading to be performed. Within this framework, we have developed a theory of creative reading. The theory is implemented in the ISAAC (Integrated Story Analysis And Creativity) system, a reading system which reads science fiction stories.

This paper reports on an exploratory study of diagrammatic reasoning. Concurrent think-aloud protocols and gestures of subjects solving a set of device behavior hypothesis problems presented as labeled diagrams were collected. In addition to analyzing verbal protocols, the gestures and marks made by the subjects were examined and used to annotate encoded verbal data. A model of diagrammatic reasoiung in this task is proposed and compared with results of analyzing the protocols. Besides lending support to results of previous experimental studies, this study also revealed some interesting aspects of diagrammatic reasoning that merit further investigation.

NTD-Soar is a model of the perceptual, cognitive, and motor actions performed by the NASA Test Director as he utilizes the materials in his surroundings and communicates with others to prepare for a Space Shuttle Launch. The model, built within the framework of a serial symbolic architecture, is based on a number of independently designed general cognitive capabilities as well as a cognitive analysis of a particular task. This paper presents a detailed description of the model and an assessment of its performance when compared to human data. NTD-Soar's ability to display human-like real-time performance demonstrates that symbolic models with a serial bottleneck can account for complex behaviors which appear to happen in parallel, simply by opportunistically interleaving small elements of the different subtasks.

Several connectionist models have been supplying non-classical explanations to the challenge of explaining systematicity, i.e.. structure sensitive processes, without merely being implementations of classical architectures. However, lately the challenge has been extended to include learning related issues. It has been claimed that when these issues are taken into account, only a restricted form of systematicity could be claimed by the connectionist models put forward so far. In this paper we investigate this issue further, and supply a model and results that satisfies even the revised challenge.

Throughout history the concept of infinity has played an important role in almost every branch of human knowledge Paradoxically, very little effort has been made by the various theoretical schools in Cognitive Science to study this fascinating aspect of human mental activity. The study of subdivision offers an interesting subject matter to address the question of how the idea of infinity in the small emerge in our minds. 32 students, aged 8, 10, 12 and 14 (high and low intellectual-academic performers), participated in this study, in which a version of one of Zeno's paradoxes was analyzed by means of individual interviews. Results suggest that between ages 10 and 12, a certain intuition of the entailments of subdivision emerges, remaining very labile afterwards and being very influenced by the context. 66 % of the 12- and 14-year-old children said that the process involved in the paradox comes to an end. Less than 25 % considered (with deep hesitations) the possibility that the process might continue endlessly. This suggests that the classic piagetian view that the indefinite subdivision is mastered at the period of formal operations must be reassessed. Some epistemological consequences based on an embodiedcognition oriented perspective are discussed.

Various cognitive and compuUtional models have addressed the use of previous experience to understand a new domain. In particular, research in case-based reasoning has explored the ideas of retrieving and adapting previous experience in the form of cases, which can only be retrieved when they are appropriately indexed. In contrast to learning new indexes, reindexing of existing cases has received little attention. The need for re-indexing a case arises when a previous situation has been incorrectly or incompletely understood. W e describe a novel approach to re-indexing which integrates results from two different areas: multiple viewpoints used in intelUgent tutoring systems and introspective questioning used in metacognitive activities. Furthermore, we apply ideas from CaseBased Reasoning to the re-indexing process itself. The revised index can be tested by active interaction with the agent's environment. An example of our implementation, JULIAN, will illustrate the re-indexing process.

Thinking is governed by abstract schemas. Verbal protocols illustrate spontaneous use, by logically unsophisticated subjects, of the schema known as modus tollens. The tollens inference schema appeared embedded within two reasoning strategies, the classical reductio ad absurdum and reasoning by elimination. The psychological reality of modus tollens is implicitly assumed by many theories in cognitive science and the hypothesis that it is a basic component of human cognition cannot be dismissed.

We developed a computational model of similarity judgment in problem-solving contexts. The model first attempts to transform an object to another using the knowledge of the domain, the strategy, and the goal. If the transformation succeeds, new feature about transformability is created. A similarity of an object to another is computed, based on the created features. If the model fails to create a new feature, it computes a similarity by feature comparison in the same way as the contrast model. An important prediction of the model is that the asymmetry of similarity judgments is caused by the directionadity of the problem-solving skills. W e examined the model's prediction. The materijJ was the Tower of Hanoi puzzle. Subjects were required to rate the similarities of one state to the goal as well as those of the goal to a state. In Experiment 1, we taught one group of subjects the 'move-pattern strategy' that induced learners to acquire highly directional skills, and compared their judgments with those by naive subjects. The asymmetry was observed only in the judgments by the trained subjects. The second experiment showed that the results of the experiment 1 could not be attributed to the 'prototypicality' of the goal.

A connectionist model was developed to investigate the relationship between global and local information in visual perception, and an experiment tested a prediction generated by the model. The research focused on the fact that processing of global information is found to dominate processing of local information in many tasks ("global precedence"). The connectionist model demonstrated that global precedencecan arise out of simple parallel processing. The experintent demonstrated that rotating global elements eliminates Global Precedence. This empirical result supports the possibility, raised by the model, that Global Precedence is due in part to simplicity of Input-Output mapping.

MacDonald, Pearlmutter, and Seidenberg (1993) propose an alternative to the dominant view in sentence processing that syntactic ambiguities are resolved by heuristics based on structural simplicity. MacDonald et al. argue that such ambiguities can be defined in terms of alternatives associated with information in individual lexical items, and thus that syntactic ambiguities can be resolved by lexical disambiguation mechanisms relying on access to the relative frequencies of alternatives and to biases created by contextual constraints. We present evidence from a computer simulation of the use of frequency-based and contextual constraints in the processing of the main verb/reduced relative syntactic ambiguity, showing that frequency and relatively limited contextual information from a sample of natural language can interact sufficiently to model basic results in the literature.

If we are to develop language processing systems that model human capabilities and performance, we must identify corespondences between the grammatical features and meaning of language and employ them in our computatiraial models of sentence interpretation. In this paper, we present a computational model of sentence interpretation and a theory of compositional semantics. Our model provides a method for addressing a range of lexical novelty (e.g., novel verbs, novel uses of known verbs), relying on a semantic representation that maintains principled correspondences with syntactic form. In our approach. syntactic structure preserves critical information about the hierarchical structure of semantic interpretations. This property of the semantic represoitation along with restrictions on semantic interpretations enable the model to infer the semantics of novel verbs, disambiguate the semantics of known verbs, and determine the contributions that verb arguments make to sentence interpretation in a constrained and principled manner. This research offers a fruitful approach for using linguistic analysis to address the recovery of meaning in natural language processing systems.

In this paper, we investigate the interaction between linguistic and ncn-linguistic processes by considering the role of functional reasoning in understanding design specifications written in natural language. W e desaibe KA , an experimental modelbased interpretation and design system which understands English language descriptions of the design problems it solves, and examine whether KA's problem-solving amiabilities help i) ascertain the relevance of ambiguous design ^lecifications and ii) identify unspecified relations between design requirements. Our results demonstrate that augmenting language processing with the ability to reason about function along the lines suggested in K A provides effective solutions to these problems in particular as well as to other problems in natural language understanding.

Several models of categorization suggest that fixed inputs (features) are combined together to create categorization rules. It is also possible that categorization influences what features are perceived and used. This experiment explored the possibility that categorization training influences how an object is decomposed into parts. In the first part of this experiment, subjects learned to categorize objects based on particular sets of line segments. Following categorization training, subjects were tested in a whole-part decomposition task, making speeded judgments of "does whole X contain probe Y." All diagnostic and nondiagnostic category parts were used as parts within the whole objects, and as probes. Categorization training in the first part of the experiment affected performance on the second task. In particular, subjects were faster to respond when the whole object contained a part that was diagnostic for categorization than when it contained a nondiagnostic part. Whe n the probe was a diagnostic category part subjects were faster to respond that it was present than absent, and when the probe was a nondiagnostic part, subjects were faster to respond that it was absent than that it was present. These results are discussed in terms of perceptual sensitivity, response bias, and the modulating influence of experience.

Systematicity, the ability to represent and process stnicturally related objects, is a significant and pervasive property of cognitive behaviour, and clearly evident in language. In the case of Connectionist models that leam from examples, systematicity is generalization over examples sharing a conmion structure. Although Connectionist models (e.g., the recurrent network and its variants) have demonstrated generalization over structured domains, there has not been a clear demonstration of strong systematicity (i.e., generalization across syntactic position). The tensor has been proposed as a way of representing structured objects, however, there has not been an effective learning mechanism (in the strongly systematic sense) to explain how these representations may be acquired. I address this issue through an analysis of tensor learning dynamics. These ideas are then implemented as the tensor-recurrent network which is shown to exhibit strong systematicity on a simple language task. Finally, it is suggested that the properties of the tensor-recurrent network that give rise to strong systematicity are analogous to the concepts of variables and types in the Classical paradigm.

Evidence suggests that an early representation in the visual processing of orthography is neither visual nor phonological, but codes abstract letter identities (ALIs) independent of case, font, size, etc. How could the visual systeai come to develop such a representation? W e propose that, because many letters look similar regardless of case, font, etc., different visual forms of the same letter tend to appear in visually similar contexts (e.g., in the same words written in different ways) and that correlationbased learning in visual cortex picks up on this similarity among contexts to produce ALIs. W e present a simple selforganizing Hebbian neural network model that illustrates how this idea could work and that produces ALIs when presented with appropriate input.

The Representation of ignorance is a long standing challenge for researchers in probability and decision theory. During the past decade, Artificial Intelligence researchers have developed a class of reasoning systems, called Truth Maintenance Systems, which are able to reason on the basis of incomplete information. In this paper we will describe a new method for dealing with partially specified probabilistic models, by extending a logic-based truth maintenance method from Boolean truth-values to probability intervals. Then we will illustrate how this method can be used to represent Bayesi&n Belief Networks — one of the best known formalisms to reason under uncertainty — thus producing a new class of Bayesian Belief Networks, caUed Ignorant Belief Networks, able to reason on the basis of partially specified prior and conditional probabilities. Finally, we will discuss how this new method relates to some theoretical intuitions and empirical findings in decision theory and cognitive science.

In this paper, we present results ftom two empirical studies in which subjects diagnosed faults that occurred in a computerbased, dynamical simulation of an oil-fired marine power plant, called Turbinia. Our results were analyzed in the framework of dual problem space search (DPSS), in which non-routine diagnosis was characterized as a process of generating hypotheses to explain the observed faults, and testing these hypotheses by conducting experiments. In the first study, we found that the less-efficient subjects conducted significantly more experiments, indicating a strong bottom-up bias in their diagnostic sU-ategy. In tiie second study, we examined the effects of imposing external resource bounds on subjects' diagnostic strategies. Results indicated that constraints on diagnosis time led to a reduction in the number of actions performed and components viewed, without appearing to affect diagnostic performance. Constraints on the number of diagnostic tests reduced search in the experiment space, which appeared to negatively affect performance. Taken together, these suggest results that subjects' diagnostic strategies were sensitive to consti-aints in tiie external task environment. We close with a sketch of how DPSS might be augmented to include effects due to external resource bounds.

In a guessing game, Ss reconstruct a sequence by guessing each successive element of the sequence from a finite set of alternatives, receiving feedback after each guess. A n upper bound on Ss knowledge of the sequence is given by H, the estimated entropy of the numbers of guesses. The method provides a measure of learning independent of material type and distractors, and the resulting data set is very rich. Here, the method is apK plied to artificial grammar learning; Ss were exposed to strings from a finite state grammar and subsequently distinguished between strings that followed or violated the grammar reliably better than Ss who had not seen the learning strings (but who themselves performed at above chance levels). Ss knowledge of the strings, H, reflected both grammaticality and exposure to learning strings, and was correlated with overall judgement performance. For non-grammatical strings, the strings that Ss knew most about were those they found most difficult to c\assify correctly. These results support the hypothesis that fragment knowledge plays an important part in artificial grammar learning, and we suggest that the guessing game paradigm is a useful tool for studies of learning and memory in general.

CELIA is a computational model of how a novice student can quickly become competent at a procedural task through observing and understanding an expert's problem solving. This model was inspired by protocol studies, and implemented in a computer program. This model of a student's effective learning suggests some implications for teaching novices in a new domain. These may be relevant for both human teaching and intelligent tutoring. The implications include: encourage the student to predict, interactive step-by-step presentation of example steps, encourage self-explanation by the student, order example steps to match their logical order, give a variety of examples in early instruction, allow flexible interaction with the student, and present bztsic background concepts prior to examples. These implications represent hypotheses that follow from the learning model; they suggest further research.

History tools play an important part in supporting human computer interaction. Most research in history tools has focussed on user interaction histories. In contrast, this paper presents a theoretical framework for artifact history and describes a computer based design environment which implements embedded artifact history. The most promising area for history tools is in collaborative design, helping users to understand others' as well as one's own previous work.

Reseaichen in AI often say that certain types of reference are baaed on perception. Their models, however, do not reflect perceptual functioning, but instead represent denotation, an intellectuaUy modeled relation, by using exact feature matching in a serial device as the basic mechanism for reference. I point out four problems in this use of denotation: substitution of an intellectual model for a perceptual one; unclarity about the nature of referential identification; relative neglect of the role of contrast in reference; and inexact matches. I then suggest an alternative theoretical account for perceptually based indexical reference, the figure-ground model, and I e:q>lain how this model handles the four problems.

When people categorize an object, they often encode a certain number of its properties for later classification. In Schyns and Murphy (in press), we suggested that the way people group objects into categories could induce the learning of new dimensions of categorization--i.e., dimensions that did not exist prior to the experience with the categorization system. In this research, we examine whether the context of known concepts can influence feature extraction. The first experiment simply tested whether the context of different object categories could change the perception of the same target stimuli. The second experiment examined whether learning category B given the concept of category A may result in different features being learned that learning A given B. The results showed that the context of known concepts influence the features people learn to represent object categories.

The production of a disyllabic word is speeded up by advance (off-line) knowledge of the first syllable, but not by knowledge about the second syllable (Meyer, 1990). By contrast, when first-syllable or second-syllable primes are presented during the production of a disyllabic word (on-line), both primes yield a facilitatory effect (Meyer & Schriefers, 1991). In this paper, the computational model of word-form encoding in speaking developed in Roelofs (1992b, submitted) is applied to these contradictory findings. Central to the model is the proposal by Levelt (1992) that morphemic representations are mapped onto stored syllable programs by serially grouping the morphemes' segments into phonological syllables, which are then used to address the programs in a syllabary. Results of computer simulations reported in this paper show that the model resolves the empirical discrepancy.

This paper presents the results of a study which examined children's ideas about the origin and differentiation of species. The focus of this paper is on the epistemic constructs associated with children's explanatory frameworks. Two groups of elementary school students, 9- year-olds and 12-year-olds, were interviewed using a semistructured questionnaire. The results indicate that most children explain the phenomena of speciation in terms of a conceptual framework that strongly resembles either early Greek or later renaissance variants of Essentialist theories in biology. Children also demonstrate a spontaneous understanding of important epistemic constructs associated with theoretical frameworks. For example, most children show an explicit awareness of the boundaries of their theoretical frameworks and have some idea of the phenomena that such a framework can and should explain. Many children treat questions about the origins of the first animal and plant species as "first questions," or questions which are in principle unanswerable. The children appear to distinguish between facts that they as individuals lack but that are probably known by experts, domain problems that are unsolved but could in principle be answered by biological theories, and problems that are beyond the explanatory scope of biological theories.

Much of the debate on rule-based vs. connectionist models in language acquisition has focussed on the English past tense. This paper investigates a new area, the acquisition of verb subcategorization. Verbs differ in how they express their arguments or subcategorize for them. For example, "She gave him a book." is good, but "She donated him a book." sounds odd. The paper describes a connectionist model for the acquisition of verb subcategorization and how it accounts for overgeneralization and learning in the absence of explicit negative evidence. It is argued that the model presents a better explanation for the transition from the initial rule-less state to final rule-like behavior for some verb classes than the symbolic account proposed by Pinker (1989).

Face recognition stands out as a singular case of object recognition: Although most faces are very much alike, people discriminate between many different faces with outstanding efficiency. Even though little is known about the mechanisms of face recognition, viewpoint dependence — a recurrent characteristic of research in face recognition — could help to understand algorithmic and representational issues. The current research tests whether learning only one view of a face could be sufficient to generalize recognition to other views of the same face. Computational and psychophysical research (Poggio & Vetter, 1992) showed that learning one view of a bilaterally symmetric object could be sufficient for its recognition, if this view allows the computation of a symmetric, "virtual," view. Faces are roughly bilaterally symmetric objects. Learning a side-view — which always has a symmetric view — should allow for better generalization performances than learning the frontal view. Two psychophysical experiments tested these predictions. Stimuli were views of shaded 3D models of laserscanned faces. The first experiment tested whether a particular view of a face was canonical. The second experiment tested which single views of a face give rise to best generalization performances. The results were compatible with the theoretical predictions of Poggio and Vetter (1992): learning a side view allows better generalization performances than learning the frontal view.

The experiment reported in this paper provides evidence that there are at least two independent implicit learning mechanisms in implicit learning: an efficiency mechanism, which underlies changes in reaction time to patterned stimuli, and a conceptual fluency mechanism, which underlies the ability to make judgments about stimuli based on implicit knowledge. Each of these implicit mechanisms is independent of explicit learning. Subjects performed a serial reaction time task under one of three learning conditions (nonattentional, attentional and observational) for one of three study lengths (2, 6 or 12 blocks). Subjects then completed five tests of their knowledge: attentional and nonattentional reaction time tasks (measuring two kinds of efficiency learning), awareness questionnaire (measuring explicit knowledge) , a generation task, and a conceptual fluency task. Correlation analyses and criterion analyses found no dependencies between the measures in low awareness subjects. In addition, the measures were influenced differently by the independent variables of learning condition and study length; these dissociations indicate separate underlying mechanisms. Implications of the existence of multiple implicit mechanisms for connectionist modeling of implicit learning are drawn.

Social agents, both human and computational, inhabiting a world containing multiple active agents, need to coordinate their activities. This is because agents share resources, and without proper coordination or "rules of the road", everybody will be interfering with the plans of others. As such, we need coordination schemes that allow agents to effectively achieve local goals without adversely affecting the problem-solving capabilities of other agents. Researchers in the field of Distributed Artificial Intelligence (DAI) have developed a variety of coordination schemes under different assumptions about agent capabilities and relationships. Whereas some of these research have been motivated by human cognitive biases, others have approached it as an engineering problem of designing the most effective coordination architecture or protocol. We propose reinforcement learning as a coordination mechanism that imposes little cognitive burden on agents. More interestingly, we show that a uniform learning mechanism suffices as a coordination mechanism in both cooperative and adversarial situations. Using an example block-pushing problem domain, we demonstrate that agents can use reinforcement learning algorithms, without explicit information sharing, to develop effective policies to coordinate their actions both with agents acting in unison and with agents acting in opposition.

The self-advertising, at least, suggests that 'situated cognition' involves the most fundamental conceptual reorganization in AI and cognitive science, even appearing to deny that cognition is to be explained by mental representations. A. Vera and H. Simon have rebutted many of these claims, but they overlook an important reading of situated arguments which may, after all, involve a genuinely revolutionary insight.

We monitored readers' eye-movements to examine the time-course of discourse and semantic influences in syntactic ambiguity resolution. Our results indicate immediate and simultaneous influences of referential context and local semantic fit in the reading of reduced relative clauses (i.e.. The horse raced past the bam fell.). These results support a model of sentence processing in which alternatives of a syntactic ambiguity are differentially activated by the bottom-up input, and syntactically-relevant contextual constraints simultaneously add activation to their supported alternatives. Competition between comparably active alternatives may then cause slowed reading times in regions of ambiguity.

In some tasks (e.g., assigning meanings to ambiguous words) humans produce multiple distinct alternatives in response to a particular stimulus, apparently mirroring the environmental probabilities associated with each alternative. For this purpose, a network architecture is needed that can produce a distribution of outcomes, and a learning algorithm is needed that can lead to the discovery of ensembles of connection weights that reproduce the environmentally specified probabilities. Stochastic symmetric networks such as Boltzmann machines and networks that use graded activations perturbed with Gaussian noise can exhibit such distributions at equilibrium, and they can be trained to match environmentally specified probabilities using Contrastive Hebbian Leaning, the generalized form of the Boltzmann Learning algorithm. Learning distributions exacts a considerable computational cost as processing time is used both in settling to equilibrium and in sampling equilibrium statistics. The work presented here examines the extent of this cost and how it may be minimized, and produces speedups of roughly a factor of 5 compared to previously published results.

Models of human parsing typically focus on explaining syntactic preferences and garden-path phenomena. This paper explores another aspect of the processing of syntactic ambiguity—the successful revision of previously preferred structure. In the competitive attachment model of parsing, a hybrid connectionist network directly represents the attachment structure among phrasal nodes in a parse tree. A syntactic ambiguity leads to a network of alternative attachments that compete for numeric activation. The winning attachments are determined within a parallel operation that simultaneously revises earlier attachments as needed when initially attaching a new phrase to the developing parse tree. Because of the unique parallel structuring operation, the competitive attachment model provides a unified explanation of human preference and recovery mechanisms in parsing. The paper demonstrates this ability by showing how the model accounts for recency effects in human syntactic processing. In the parsing network, a mechanism of decay, which is independently needed to manage the finite pool of processing nodes, allows more recent phrases to compete more effectively than less recent phrases for new attachments. The effect of decay on the attachment competition underlies a unified account of psycholinguistic observations of recency, both in initial syntactic preferences and in the revision of erroneous attachments.

Past research in cognitive science reveals that prototypical configurations of domain objects, called perceptual chunks, underlie the abilities of experts to solve problems efficiently. Little research, however, has been carried out on the mechanism used for learning perceptual chunks from solving problems. The present paper addresses this issue in the domain of geometry proof problem-solving. We have developed a computational model that chunks, from problem diagrams, configuration of the elements which are visually grouped together, based on perceptual chunking criterion. This criterion, called recognition rules, reflects how people see problem diagrams and thus works effectively to determine which portion of problem diagrams are more likely to be grouped as a chunk. This distinguishes the proposed method from the goal- oriented chunking techniques used in machine-learning community. Experiments on solving geometry problems show that our technique can detect essential diagram configurations common to many problems. Additionally, implications of the recognition rules are discussed from a cognitive point of view.

Much problem solving and learning research in math and science has focused on formal representations. Recently researchers have documented the use of unschooled strategies for solving daily problems -- informal strategies which can be as effective, and sometimes as sophisticated, as school-taught formalisms. Our research focuses on how formal and informal strategies interact in the process of doing and learning mathematics. We found that combining informal and formal strategies is more effective than single strategies. We provide a theoretical account of this multiple strategy effect and have begun to formulate this theory in an ACT-R computer model. We show why students may reach common impasses in the use of written algebra, and how subsequent or concurrent use of informal strategies leads to better problemsolving performance. Formal strategies facilitate computation because of their abstract and syntactic nature; however, abstraction can lead to nonsensical interpretations and conceptual errors. Reapplying the formal strategy will not repair such errors; switching to an informal one may. We explain the multiple strategy effect as a complementary relationship between the computational efficiency of formal strategies and the sense-making function of informal strategies.

This research aims to clarify, by constructing and testing a computer simulation, the use of multiple representations in problem solving, focusing on the role of visual representations. We model the behavior of an economics expert as he teaches some economics principles while drawing a graph on a blackboard. Concurrent verbal protocols are used to guide construction of a production system. The model employs representation-specific data structures and rules. The graph on the blackboard is represented by a bit map; the pictorial working memory (WM ) and long term memory (LTM) representations are node-link structures of a pictorial nature; the auditory WM and LTM representations are node-link structures of a verbal-semantic nature. Pieces from the different representations are linked together on a sequential and temporary basis to form a reasoning and inferencing chain, using cues from LTM and from the external graph. The expert used two representations so as to exploit the unique advantages of each. The graphical representation served as a place holder during reasoning, as well as a summary. The verbal-semantic representation served to give semantic meaning and causal background. Both could initiate reasoning chains. We compare the expert's behavior with novices' trying to learn the same principles.

Three principles governing the operation of the lexical pathway in a model of reading single words aloud were applied to the question of learning, as measured by times to initiate correct pronunciations. I. At the lexical level, a target word activates a neighborhood of orthographically similar entries in the lexicon. II. At the phoneme level, the correct phonemes in the phonemic spelling of the word compete with the other active phonemes. III. At the naming level, the pronunciation is composed of a conjunction of phonemes. These principles were tested using the dau from a 4-year-old beginning reader (U) , resulting in a goodness-of-fit R^2 = .44. When a rule pathway using grapheme-phoneme correspondences was added to the lexical pathway, the goodness-of-Ht was comparable (R^2 = .46). When single entries were accessed along the lexical pathway, instead of word neighborhoods, and grapheme-phoneme conespondences were accessed along the rule pathway, as in standard dual-route models, the goodness-of-fit R^ 2 fell to .27. Although the modelfltting supported the importance of neighborhood activation and failed to support the importance of rules. grapheme-phoneme correspondences were overtly used by LT in the initial trials with words and when feedback indicated an errorful pronunciation. Thus, rule application may be relatively slow in normal fluent word naming, but may still play a strategic role in attempts to initially decode letter strings or to correct errors.

Many efforts have been made to use nonnative theories of rational decision-making, such as Bayesian decision theory, to consruct and model agents exhibiting intelligent behavior. In order to accommodate agents possessing only limited computational resources to apply to their decision making, however, a significant change is required in how the role of formal rationality is to be viewed. This paper argues that rationality is best seen as a property of the relationship between the agent and a designer. Such a perspective has several consequences for the design and modelling of agents, bearing on assessment of rationality, induction, reactivity, and metalevel control. It also illuminates several concerns put forth by critics of the work of the artificial intelligence community.

Models of rationality typically rely on underlying logics that allow simulated agents to entertain beliefs about one another to any depth of nesting. We argue that representations of individual deeply nested beliefs are in principle unnecessary for any cooperative dialogue. We describe a simulation of such dialogues in a simple domain, and attempt to generalize the principles of this simulation, first to explain features of human dialogue in this domain, then those of cooperative dialogues in general. We propose that for the purposes of cooperative interaction, the status of all deeply-nested beliefs about each concept can be conjoined into a single represented value, which will be Affected by reasoning that might be expected to lead to conclusions in terms of deeply-nested beliefe. We concede that people are capable of using individual deeply-nested beliefs to some degree, but such beliefs need only be handled explicitly in dialogues involving secrecy or deception.

Previous work in visual cognition has extensively explored the power of parts-based representations of objects for recognition, categorization, and functional reasoning. We propose a novel, parts-based representation of objects, where the parts of an object are found by grouping together object elements that move together over a set of images. The distribution of object configurations is then succinctly described in terms of these functional parts and an orthogonal set of modal transformations of these parts. If the distribution has a natural set of principal axes, the computed modes are stable and functionally significant. Moreover, the representation is always unique and robustly computable because it does not rely critically on the properties of any particular element in any particular instance of the object. Most importantly, the representation provides a set of direct cues to object functionality without making any assumptions about object geometry or invoking any high-level domain knowledge. This robustness and functional transparency may be contrasted with standard representations based on geometric parts, such as generalized cylinders (Marr and Nishihara, 1978) or geons (Biederman, 1987), which are sensitive to accidental ahgnments and occlusions (Biederman, 1987), and which only support functional reasoning in conjunction with high-level domain knowledge (Tversky and Hemenway, 1984).

Synchronous firing of neural units has recently been proposed as a new way of solving the variable binding problem in connectionist networks. Firing synchrony appears to be unrelated to earlier methods of variable binding, nearly all of which can be analyzed as species of tensor product representations, where vectors representing variables and values are bound together with the outer product. In this paper, we argue that, despite appearances, firing synchrony is also a case of tensor product representation. This analysis exposes two logically independent components of the synchronous firing idea. The most obvious is the idea of using time as a resource: spatio-temporal patterns of activation are used. This, we argue, is a purely implementational issue which does not bear on the complexity issues of variable binding. In contrast, the second idea does bear on genuinely representational issues, and is the source of most of the formal properties claimed for the synchrony scheme. Rather than explicitly binding a semantic role like giver to a semantic filler like John, these two are implicitly bound—by explicitly binding each to a common formal role, via the tensor product. The analysis situates synchronous firing in a typology of alternative variable binding schemes.

The perceptual principles that allow people to group visually similar objects into entities, or groups, have been called the Gestalt Laws of perception. Two well known principles of perceptual grouping are proximity and similarity: objects that lie close together are perceived to fall into groups; objects of similar shape, size or color are more likely to form groups than objects differing along these dimensions. While the primary function of these "laws" is to help us perceive the world, they also enter into our communications. People can build on assumptions about each other's perception of the world as a basis for simplifying discourse: for example, we invariably refer to collections of objects simply by gesturing in their direction and uttering "those." The current work describes an algorithm that simulates parts of the visual grouping mechanism at the object level. The system uses feature spaces and simple ranking methods to produce object groupings. Computational aspects of this system are described in detail and its uses for enhancing multi-modal interfaces are explained.

When agents cooperate to solve complex problems in the realworld, they must choose which information to communicate from the mass of information that might affect the problem. A speaker should communicate the information that will be most helpful to the other agent. However, the speaker may not have a great deal of knowledge about the other. In addition, the speaker is also involved in reasoning about the collaborative problem solving task. So, processing that is done solely to select information will be taken from the resources available to work on the primary problem. In this paper we present preliminary work on a new approach to selecting information that should be included in a dialogue. Our approach uses the speaker's knowledge of its own problem solving to determine how useful some piece of information might be to other agents. Consequently, the speaker can make its decision to include information in the dialogue using no additional knowledge and few additional computational resources beyond those required to reason about the primary problem solving task. We suggest heuristics which translate problem solving into estimates of how useful information will be for others.

Handling unanticipated events during problem solving is difficult enough when an agent is operating by itself. When the agent is part of a cooperative distributed problem solving (CDPS) system, the task's difficulty increases dramatically. Now the agent is forced to consider the effect of the event not only on itself, but also on others and the group as a whole. It must also consider who should handle the event and the likely impact that actions taken to diagnose the event or respond to it may have on other agents. In this paper, we discuss preliminary work aimed at developing a process for handling events during multiagent cooperative problem solving. The domain in which the work is being done is cooperating multiple autonomous underwater vehicles (AUVs). However, the approach should have broader applicability to almost any realworld cooperative problem solving task involving autonomous or nearly autonomous agents.

This paper describes a prototype of a simulated physics student that learns by interacting with a human tutor. The system solves physics problems while showing its work on a workstation screen, and the tutor can intervene at certain points during problem-solving to advise the simulated student This prototype constitutes an initial cognitive task analysis of the skill of learning from a tutor, which prescribes several tutoring practices that appear to be plausible for both human and computer tutors.

A previously constructed prototype expert system was extended to include case-based reasoning/learning, in order to determine if the automated problem solving behavior could be improved. The initial expert system was developed by using an inductive machine learning technique on 9,445 data records of pregnant women, providing production rules to predict preterm delivery. Its predictive accuracy was tested on a separate set of 9,445 data records. Next, the capability to reason from both production rules and input test cases was added to the system, in addition to the capability to internally modify its confidence in each piece of knowledge (rule or case) and the relative importance of patient attributes which appear to be predictive of preterm delivery. The system was structured such that the accuracy of either type of reasoning could be measured individually to determine how rule-based and case-based reasoning perform alone, and to determine how they perform together. Results show that the predictive accuracy of the system was improved, with different trends emerging, dependent on the bias of the learning data. Neither system performed as well alone as did both together.

This paper challenges the widely accepted claim that "classical" cognitive architectures can explain the systematicity of cognition (Fodor & Pylyshyn, 1988). There are plausible ways of rendering more precise the systematicity hypothesis (as standardly formulated) in which it is entailed by classical architectures, and other plausible ways in which it is not. Therefore, it is not a determinate issue whether systematicity is entailed, and hence explained, by classical architectures. The general argument is illustrated in a particular domain, the systematicity of deductive inference. In the case of the capacity to carry out the inference modus tollens, the systematicity hypothesis can be made precise in two ways, one entailed by classical architectures, another which is not. Further, the latter, but not the former, accurately describes the actual empirical phenomenon. Put another way, the clumps that these deductive inference capacities come in are not the clumps that are entailed by classical architectures. Therefore, in this area at least, systematicity considerations count against the classical conception of cognitive architecture.

Metaphor is an elegant, concise, often startling communicative form which is employed by a speaker as a means of conveying a state of affairs to a hearer; as such, it deserves to be analysed as a speech-act, with a particular illocutionary intent and perlocutionary effect. This paper describes a hybrid symbolic/connectionist model of meuphor (SAPPER by Veale & Keane. 1993), which incorporates elements of the belief ascription model of (Wilks. Bamden & Wang, 1991). This extended framework provides a suitable computational envirormient for analysing the illocutionary intent of the speaker, and perlocutionary effect upon the hearer's belief space, of a broad class of metaphors with an observable ameliorative/pejorative connotation.

Theories of skill acquisition have made radically different predictions about the role of means-ends analysis in acquiring general rules that promote effective transfer to new problems. Under one view, means-ends analysis is assumed to provide the basis for efficient knowledge compilation (Anderson, 1987), whereas under the alternative view means-ends analysis is believed to disrupt rule induction (Sweller, 1988). We suggest that in the absence of a specific goal people are more likely to use a rule-induction learning strategy, whereas providing a speciflc goal fosters use of means ends analysis, which is a non-rule-induction strategy. We performed an experiment to investigate the impact of goal specificity and systematicity of rule-induction strategies in learning and transfer within a complex dynamic system. Subjects who were provided with a specific goal were able to solve the initial problem, but were impaired on a transfer test using a similar problem with a different goal, relative to subjects who were encouraged to use a systematic rule-induction strategy to freely explore the problem space. Our results support Sweller's proposal that means-ends analysis leads to specific knowledge of an isolated solution path, but docs not provide an effective method for learning the overall structure of a problem space.

A model based on coupled mechanisms for place recognition, path integration, and maintenance of head direction in rodents replicates a variety of neurophysiological and behavioral data. Here we consider a task described in [Collett et al. 1986] in which gerbils were trained to find food equidistant from three identical landmarks arranged in an equilateral triangle. In probe trials with various manipulations of the landmark array, the model produces behaviors similar to those of the animals. We discuss computer simulations and an implementation of portions of the model on a mobile robot.

The emerging function of verb inflections in German lenguage acquisition is modeled with a connectionist network. A network that is initially presented only with a semantic representation of sentences uses the inflectional verb ending -t to mark those sentences that are low in transitivity, whereas all other verb endings occur randomly. This behavior matches an early stage in German language acquisition where verb endings encode a similar semantic rather than a grammatical function. When information about the siuface structure of the sentence is added to the input data, the network learns to use the correct verb inflections in a process very similar to children's learning. This second phase is facilitated by the semeintic phase, suggesting that there is no shift from semantic to grammatical encoding, but rather an extension of the initial semantic encoding to include grammatical information. This can be seen as evidence for the strong version of the functionalist hypothesis of language acquisition.

An unexplored means by which analogical transfer might take place is through indirect priming through the interaction of text comprehension and memory retrieval processes. Remind is a structured spreading-activation model of language understanding and reminding in which simple transfer can result from indirect priming from previously processed source analogs. This paper describes two experiments based on Remind's priming-based transfer framework. In Experiment 1, subjects (1) summarized analogous source stories' common plot; (2) rated the comprehensibility of targets related to sources by similar themes, contexts, or themes and contexts; then (3) described any sources incidentally recalled during target rating. Source/target similarity influenced comprehensibility and reminding without any explicit mapping or problem-solving. In Experiment 2, subjects (1) rated each story's comprehensibility in source/target pairs having similar relationships to each other as in Experiment 1; then (2) rated source/target similarity. Analogous targets were rated as more comprehensible than non-analogous targets. Both experiments imply that transfer can be caused by activation of abstract knowledge representations without explicit mapping.

Creative designers often see solutions to pending design problems in the everyday objects surrounding them. This can often lead to innovation and insight, sometimes revealing new functions and purposes for common design pieces in the process. We are interested in modeling serendipitous recognition of solutions to pending problems in the context of creative mechanical design. This paper characterizes this ability, analyzing observations we have made of it, and placing it in the context of other forms of recognition. We propose a computational model to capture and explore serendipitous recognition which is based on ideas from reconstructive dynamic memory and situation assessment in case-based reasoning.

We argue that discourse plans must capture the intended causal and decompositional relations between communicative actions. W e present a planning algorithm, DPOCL , that builds plan structures that property capture these relations, and show how these structures are used to solve the problems that plagued previous discourse planners, and allow a system to participate effectively and flexibly in an ongoing dialogue.

Most graphic and tabular displays are relational information displays—displays that represent relational information, which is a relation on a set of dimensions. In this paper, we argue that relational information displays are distributed representations—representations that are distributed cross the internal mind and the external environment, and display-based tasks are distributed cognitive tasks—tasks that require the interwoven processing of internal and external information. The basic components of relational information displays are dimensions. Through a theoretical analysis of dimensional representations, we identified four major factors that affect the representational efficiencies of relational information displays: the distributed representation of scale information, the relation between psychological and physical measurements, the interaction between dimensions, and the visual and spatial properties of dimensions. Based on the representational analysis of relational information displays, we proposed a representational taxonomy of relational information displays. This taxonomy can classify most types of relational information displays. In addition, it can be used as a theoretical framework to study the empirical issues of relational information displays in a systematic way.

The focus of this symposium will be on the relevance of animal cognition to human cognition. The speakers at the Symposium will present research in animal cognition and will discuss its relevance to human cognitive systems. After their presentations, the speakers and the audience will participate in a moderated discussion of the relevance of animal cognition to human cognition and to cognitive science as a whole.

The notion of cases arises in various guises in a number of areas of research within Cognitive Science. Recent theories of situated cognition, for example, have argued that learning occurs most felicitously in circumstances that most closely resemble those of eventual use (Brown, Collins, & Duguid, 1989; Norman, 1993). Furthermore, research in analogical problem-solving has shown that transfer is improved when the similarity between training problems and target problems is increased (Gick & Holyoak, 1980). Both positions, therefore, support the argument that instruction based on cases is more likely to be usefully applied in practice than instruction based strictly on abstracted principles. Others have gone further to argue that learning through exposure to real cases is not only beneficial, but essential for attaining expertise. For example, Dreyfus has asserted (Dreyfus & Dreyfus 1986) that advanced stages of expertise can only be achieved through practice with a large number of cases. Moreover, advocates of Case-Based Reasoning (Kolodner, 1994) have argued that the process of acquiring expertise is really one of accumulating experience with a succession of real cases and properly indexing these experiences for later retrieved. The purpose of this symposium will be to determine to what degree these views are compatible and to what degree they diverge. The presenters will endeavor to address the following questions from their own disciplinary perspectives: What is a case? How is it represented in memory? How are appropriate cases retrieved for later use? Does expertise consist (strictly) in the acquisition of a collection of past solved problems? What role should the study of cases play in the acquisition of expertise? Should they precede or follow the study of abstracted principles?

The symposium on "Visual Reasoning in Discovery, Instruction and Problem Solving" will consist of three talks focusing on the role of visual reasoning in higher level cognitive processes. As this is a newly emerging research area spanning cognitive science and artificiaJ intelligence, the symposium is designed both to inform and to stimulate interest, discussion, and further enquiry. The speakers will consider visual reasoning in three areas: scientific reasoning and discovery, learning and instruction, and problem solving. The talks will show how several different types of data can contribute to a clearer understanding of processes, mechanisms, and strategies underlying visual reasoning. First, N2uicy Nersessian will cover the discovery aspect by providing a historical view on visual representation in creative scientific reasoiung. Next, Rogers Hall will cover the educational aspect by laying out a set of genered educational questions concerning the role of representational forms, and discussing studies of how people coordinate representational resources while working on problems in different instructional and work settings. Finally, Mary Hegarty and Hari Narayanan will together cover the problem solving cispect from both experimental and computational perspectives.

This symposium will discuss various kinds of collaborative knowledge, i.e. knowledge that develops as the result of the cooperative work of groups of people.