Go to Semiotic Review of Books Home Page
Go to SRB Highlights
Go to SRB Archives


SRB

SRB Archives

This article appeared in Volume 3 (1) of The Semiotic Review of Books.

Metaphors of Cognitive Science

by Michael E. J. Masson

Foundations of Cognitive Science. Edited by Michael I. Posner/Cambridge, MA: MIT Press. 1989. 887pp. ISBN 0-262-16112-5

The goal of this volume of collected works, according to editor Michael Posner, is to lay out the basic ideas of cognitive science particularly for the benefit of newcomers to the field. Toward this end an outstanding group of contributors developed a set of chapters that explore a core set of domains centrally relevant to cognitive science. The book's opening chapter serves to set the stage for consideration of a computational approach to understanding how natural intelligence functions, with emphasis on the development of artificial analogues. This introduction is followed by three sections, one devoted to a set of chapters that articulate the fundamental methods used in exploring cognitive science, another in which the chapters describe issues of current importance in a variety of specific domains, and a final pair of chapters that consider an assessment of the field. In this review I provide a brief description of each chapter in order to convey the flavour of the book. A critical evaluation of the book follows in which issues concerning the enterprise of cognitive science and the potential role of Posner's book in that enterprise are raised.

Highlights

In the introductory chapter Simon and Kaplan advance the view that cognitive science is based on a consensus that the study of intelligence and its computational processes in humans (and other animals species), in computers, and in the abstract, have much in common. The primary contributing disciplines from this perspective are identified as psychology, artificial intelligence, linguistics, philosophy, and neuroscience. Simon and Kaplan also discuss the architecture of intelligent systems and offer as the standard model the traditional information processing architecture with its distinction between short term and long-term memory systems. Variations on this theme are used to set the stage for network and symbolic models (developed more fully in later chapters) which are seen as instantiations of parallel and serial processing, respectively. Some interesting insights are provided concerning two approaches to the study of cognitive science, reasoning and search. The division between these approaches is seen as a factor in the clustering of various chapters in the volume, with those covering language and semantics on one side and those dealing with problem solving and categorization on the other An important question is how these two lines of inquiry might be related to one another.

The section on foundations consists of a set of seven chapters that present methods of exploring issues in cognitive science, with an emphasis on computational methods. The opening chapter by Pylyshyn is devoted to the general issue of computation. He advocates the idea that cognition is a species of computing instantiated in a biological mechanism. More specifically, Pylyshyn claims that the properties of symbolic computers embody the properties of cognition and suggests that the appropriate level of comparison between computer models and cognitive processes corresponds to that of a system's architecture. The architecture takes on this insignificance because it is responsible for executing algorithms that solve particular problems. Achieving the goal of establishing equivalence between a human and a model system, then, depends on demonstrating correspondence between their respective architectures. A critical point made by Pylyshyn is that cognitive phenomena that are due to features of system architecture are cognitively impenetrable, that is, not sensitive to goals and beliefs. Phenomena that are affected by goals or beliefs may not be a direct reflection of constraints imposed by the architecture but instead may be a function of representations and processes that operate on the architecture. Therefore, in order to establish equivalence at the architectural level, it is necessary to identify phenomena that are cognitively impenetrable.

Two other chapters in the foundations section describe mutually contrasting approaches to computational methodology. The chapter by Newell, Rosenbloom and Laird puts forward the case of symbolic architectures as models of cognitions, assigning fundamental importance to symbol manipulation as the basis for cognitive processing. Aspects of the architecture show through in many features of behaviour, including reaction time, errors, and failures of rationality in decision making. The importance of the concept of architecture is in providing a total system context in which various mechanisms operate. It is a support system capable of universal computation. Newell et al. point out a potentially serious deficiency in the cognitive science enterprise when they suggest that contributors to the field rarely develop experience working on complete cognitive architectures. An alternative architecture is described in the chapter by Rumelhart who provides an elegant introduction to the fundamental principles of connectionist systems. Once again the theme of architectural influences on which algorithms or programs can be run by a system is brought forward. The key feature of connectionist systems, however, is the representation of knowledge in the connection weights that link simple processing units rather that in symbols. Connectionist systems are particularly well suited to applications, such as pattern recognition and learning, where there are interacting knowledge sources that define the constraints on task performance.

Chapters by Wasow and by Barwise and Etchemendy explore grammatical theory and semantics, respectively. Wasow outlines the important contribution of the study of generative grammar to the development of rule-based systems as the foundation for computational models of language. A further role has been to conceive of the mind as a set of specialized faculties with biologically determined properties. Despite these achievements there is a clear sensitivity to some of the key limitations of this approach to the study of cognitive science. Chief among them is the emphasis on linguistic competence in an ideal situation, at the expense of considering constraints imposed by the imperfections of the cognitive system that manifests e competence through performance. Barwise and Etchemendy examine issues regarding the relationship between language and the outside world that language is intended to represent, and in particular the relationship between language and thought. They emphasis extensional model-theoretic semantics and its role in specifying the truth conditions of sentences in a particular target language. In addition, some currently pursued alternative such as possible world semantics and situation semantics are considered.

The foundations section closes with chapters on experimental methods in psychology and neuroscience. Bower and Clapper provide a basic introduction to methods and issues in cognitive psychology, including a primer on experimental design. In an interesting contrast to Simon and Kaplan, they express a preference for quantitatively dependent variables (eg. proportion of responses of a given type) over verbal protocols. They are in league with Pylyshyn, however in their advocacy of a focus on nonstrategic factors in the study of basic properties of the cognitive system. A number of classic tool~ in the study of cognitive processes are reviewed and their application illustrated by examples drawn from research on learning, memory, and language processes. The relationship between brain and cognition is examined by Selnowski and Churchland who provide a basic review of neural structure and techniques for studying neurological processes. The stated goal of this enterprise is to develop valid neural constraints on cogitative mechanisms. One encouraging conclusion they reach is that many features of neural systems correspond to connectionist networks. At a more general level they propose that the goal of cognitive neuroscience is to produce a reductive integration of psychology and neural science in order to reveal the mind-brain relationship.

The section on domains of cognitive science consists of 11 chapters, clustered into groups on language, thinking, vision, memory, and action. The language chapters emphases acquisition (Pinker), reading (Grosz, and Rayner), and discourse (Grosz, Pollack, and Sidner). Pinker views the exploration of language acquisition as a means of providing answers to many central issues in cognitive science, such as modularity, the relationships between language and thought and between learning and innateness. A critical aspect of language acquisition as a touchstone for cognitive science is Its interdisciplinary nature, particularly involving linguistics, psychology, and computational models. An example of this interaction is Pinker's analysis of transformation of verbs into past tense. Shortcomings of the Rumelhart and McClelland (1986) connectionist account of learning this transformation are reviewed and evidence is marshalled to support the conclusion that it is not a change in environmental input (e.g. frequency of encountering irregular verbs) that leads a child out of the over-regularization stage, but apparently a maturational change in the child's brain. Pollatsek and Rayner review some basic empirical results and models from the literature on word identification and eye movement patterns in text reading. Results of research on context effects in word identification are shown to provide only equivocal support for the concept of modularity. In an effort to make contact with the notion of computational models, simulations of reading processes are briefly considered. The conclusion is that although lower level processes such as word identification may be amenable to such actively, there should be scepticism regarding the possibility of simulating complex aspects such as comprehension because of our lack of understanding of discourse processes. Grosz and her colleagues attempt to provide some headway in this regard in their chapter on discourse. They provide a summary of early work in artificial intelligence on discourse, and discuss issues concerning discourse structure and its effect on phrase level phenomena. A critical concept in this work is that coherent discourse can be construed as means of acting on ones world, and therefore analysis of discourse demands a method of recognizing plans embodied by discourse. This approach to discourse analysis places it squarely in the domain of artificial intelligence approaches to planning.

Three chapters comprise the segment on thinking. Johnson-Laird exercises the concept of mental models as a testament to the importance of the mind as a symbolic system. In this view, the basis of understanding a functional system is assumed to rest on an internal representation of a working model of the system. Johnson-Laird makes a strong case for the dim that however mental models are built, through perception, discourse, or analogy, their structure is of a kind and reasoning operates on them in familiar ways. Concepts and induction are discussed by Smith, who emphasizes prototypes as the essential representational form of concepts. A major reason for favouring prototypes is that they can be instantiated as schema representations which are amenable to the computations that humans are assumed to execute, such as inductive inferences, computation of similarity, and formation of composite concepts. An especially interesting aspect of Smith's proposal is that categorization may be viewed as a form of induction, and it is shown that heuristics involved in estimation of probabilities are similar to those used in solving the categorization problem. This analysis permits a new assessment of the reasons for deviations between estimated probabilities and solutions dictated by formal probability theory VanLehn combines issues in problem solving and skill acquisition in his discussion of experience and the distinction between knowledge rich and knowledge-lean problem domains. Although there has been satisfactory progress in developing accounts of knowledge Jean domains, VanLehn's assessment is that it has been difficult to develop accounts of knowledge acquisition that are adequate to capture the transition from novice to expert in knowledge rich domains. In addition to these observations on theory development a thought-provoking summary of empirical findings is also provided. For example, it is pointed out that practice effects have been difficult to account for in current models because the power function relating completion times with amount of practice is not a natural consequence of the computational learning mechanisms applied so fac nor do these practice effects ever stop.

The computational study of vision is reviewed by Hildreth and Ullman, who follow Marr's (1982) proposal for distinguishing three levels of describing problems in vision: computational theory algorithm, and mechanism. This separation is used to motivate the chapters emphasis on computational theory as R applies to low level visual problems such as edge detection, binocular vision, and motion, and as e applies to high-level tasks such as recognizing objects. A critical distinction drawn between low- and high level vision is that only the latter is goal directed, spatially focused, and knowledge dependent. The notion of spatially focused processing is elaborated by Allport's chapter on visual attention. Allport is successful in accomplishing two critical tasks in this chapter. First, he advocates a shift away from the classic idea that selective visual attention arises from limited capacity processing, toward the view that selected attention is critical for establishing and redirecting attention for the purpose of controlling action. Second, evidence from neuroscience is adduced to motivate this shift and implications for computational modelling of these processes, particularly within connectionist systems are considered.

Schacter provides a general introduction to historical and current issues in memory with emphasis, much as in the Bower and Clapper chapter on basic experimental psychology In addition, the concept of interdisciplinary approaches to the study of memory is well situated in a discussion of neuropsychological methods that provide interesting support for and constraints on memory theory. An evolutionary perspective is also considered and used to argue for the existence of separable memory systems. Computational theories are considered only briefly, but we are reminded of important examples of the contribution of these approaches to theory development, such as Winograd's (1975) distinction between procedural and declarative memory and the computational instantiation of schema theory as scripts or frames.

In the contribution to a two-chapter segment on action and motor control Jordan and Rosenbaum emphases a computational modelling as an approach to explaining how the cognitive system manages the problem of selecting and generating appropriate actions from a multitude of possibilities. Connectionist modelling schemes are emphasized and two methods of reducing computational load, analyzing skilled actions as a sequence of movements and sensory-motor learning, are outlined. Bizzi and Mussa-Ivaldi propose a different approach to solving the problem of motor control emphasizing the geometric and mechanical aspects of muscles that generate torque at the joints to enable movement. This approach represents a shift away from the traditional emphasis on the processing of neural signals.

In the final two chapters an assessment of cognitive science is provided from two different viewpoints. D'Andrade discusses the contribution of anthropology to the understanding of symbolic representation in various areas such as language, taxonomical emotion, and colon An especially important contribution is the notion a cultural models, when shared by a social group. D'Andrade illustrates how cultural models affect intelligent behaviour using the domain of reasoning. In contrast to the computational view, in which reasoning is based in a formal system of logic, he presents a compelling case for the proposition that reasoning is based on the application of cultural models rather than on the manipulation of abstract symbols. A critical component in this argument is that performance on versions of the Wason selection task varies across cultures as a function of the cultural relevance of the cover story. In this chapter we see some interesting challenges issued to proponents of some of the classic perspectives on cognitive science -- a point which is elaborated below. The philosophical roots of cognitive science are unearthed and re-examined by Harman. We are reminded of the ages old and fundamentally important mind-body problem -- how physiological processes give rise to changes in consciousness and belief -- and whether we can expect ever to develop a satisfactory solution. Three proposals are considered, all of which identify mental states as behavioral dispositions that arise from physical events. These ideas share an approach in which mental states are glimpsed from the outside and miss the aspects of mental life that can only be understood from a subjective point of view. Harman captures the essence of this missing point of view in the haunting conclusion that however successful these approaches to cognitive science turn out to be, they will not be adequate to "tell a blind person what it is like to see something red" (p. 843)

Reflections

As is apparent from the summary, Foundations provides a broad vista from which to view the discipline of cognitive science. It is clear that the book can serve as a valuable introduction to many of the foundations and domains of cognitive science and as an update for those already deeply committed to the enterprise. Moreover the scope and depth of the book will make it a landmark for the discipline for years to come. In the short summary of the chapters I have tried to capture a sense of this perspective by highlighting some of the major contributions of each. Given this background, I turn to a consideration of two other issues: some shortcomings of the book, and few obsenations regarding interesting aspects of the "big picture" that one cannot help but carry away after reading this volume.

By raising a few criticisms I hope to encourage improvements in later efforts of this kind and to help consumers of this literature maintain a standard of expectations that is faithful to the major objective of the cognitive science movement. It is widely accepted that this objective is to foster and exemplify the interdisciplinary study of intelligent behaviour. On one view of this objective, it is necessary to familiarize interested parties with the fundamental tools used in the discipline and to provide an awareness of the key issues in each domian. The book succeeds remarkably from this prespective. From the carefully argued proposal that machine and human intelligence are of a kind, to the masterful tutorials on the basics of experimental psychology linguistics, and the neuroscience, the elements are richly arrayed for us.

A different, fundamentally important view is not quite so well served by this volume. In this view, one needs more than an acquaintanceship or even expertise with the elemental tools in order to get into the game. One must develop skill at using the tools together to produce ideas and demonstrations that could not be accomplished within a single domain. It is certainly true that a number of the chapters provide striking examples of this integral approach to studying cognition. Some of the most impressive examples include Pinker's demonstration of the use of developmental linguistics in the construction and assessment of computational models of language acquisition, Allport's and Schacter's integration of neuroscience and experimental psychology in their respective formulations of new ideas in the study of attention and memory and D'Andrade's assessment of cultural models, their function in an evolutionary analysis of intelligence, and implications for viewing cognitive phenomena from this perspective. These chapters have a particularly significant impact because they illustrate the product and process of an interdisciplinary enterprise.

Even in those instances where domains have been integrated successfully, however there is a clear indication of how constrained these endeavors have been. In most cases an integration of only two approaches is recounted (e.g the Grosz et. al. discussion of artificial intelligence approaches to planning as a means of analyzing discourse) and when three domains are considered, one typically is given a minor role (e.g., computational models are treated only supreficially by Allport and by Schacter). This observation might be taken as evidence of how difficult is the enterprise of interdisciplinary study of cognitive science, or for that matter, any science. It may imply be too demanding to attempt an integration of three or more disciplines.

I would like to suggest that the modal tendency to combine domains in pairs is also indicative of a useful heuristic for coping with complex problems. Specifically, we can view many interdisciplinary programs as outgrowths of well cultivated metaphors that are used to promote understanding within a single specific domain. A commonly used strategy in efforts to understand and communicate about complex concepts is to do so in the context of a familiar metaphor (Ortony, 1979). Roediger (1980) provides a clear example of this for the field of memory, and the process is explored further in Johnson-Laird's chapter on reasoning with mental models. Particular examples from Foundations are easy to find. Hildreth and Uliman, for instance, advocate construction of computational models of vision as a means of exploring algorithms used by biological systems. If one understands computational techniques very well and believes that computational models are good metaphors for biological systems (a major tenet of cognitive science), then building such models is a justifiable and potentially fruitful pursuit.

It is necessary, however, to recognize the drawbacks associated with the heavy dependency on metaphor that appears to drive cognitive science. One concern is that by moving too deeply into a metaphor we become so far removed from the original phenomenon of interest that the constructs we develop begin to lose contact with it, and we can become sidetracked by the peculiarities of the metaphor itself. In addition, it is all too easy to become dependent on a single dominant metaphor. In cognitive science the symbolic manipulation metaphor clearly is preeminent, but we should welcome the application of alternative metaphors to encourage new ways of thinking. The relatively recent entry of neuroscience and neural interaction as bases for an alternative metaphor has proven every fruitful and a clear competitor in the form of connectionist models of computation has emerged.

Yet another metaphor has begun to shoulder its way into the group, although this one was not given much attention in Foundations. The fact that cognitive systems evolved according to a set of well established principles can be used as a guide to exploring the properties of intelligence (e.g. Cosmides, 1989). Schacter contributes some examples in the context of memory, but it is in D'Andrade's chapter that this approach is most forcefully demonstrated. He points out that people are poor at manipulating abstract symbols or searching decision trees, but good at learning complex conceptual structures (cultural models). It is ironic, according to D'Andrade, that when cognitive scientists began to study reasoning they used highly specialized symbol manipulation tasks (e.g., logic, chess) that people do not do well or frequently. He argues that rather than translating connectives such as if-then into truth-table relations, people understand them in terms of culturally determined temporal and causal contingencies that involve propositional content. In this view, further progress in modelling intelligence may be achieved by making use of knowledge regarding how humans solved the problem of becoming intelligent -- they developed culture.

Seen as an enterprise of metaphor cultivation, cognitive science may be judged as having mined only a small number of potential sources. As we begin to recognize this fact, and as new domains enter the field we should expect even more metaphors to surface, helping to deepen our understanding of intelligence.

References

Cosmides, Leda (1989) "The logic of social exchange: Has natural selection shaped how humans reason?" Studies with the Wason selection task. Cognition 31.1: 187-276.

Ortony, Andrew (ed.) (1979) Metaphor and Thought. Cambridge: Cambridge University Press.

Roediger, Henry L. (1980) "Memory metaphors in cognitive psychology." Memory & Cognition 8.3: 231-246.

Rumelhart, David E. and James L. McClelland (1986) "On learning the past tenses of English verbs." In Parallel Distributed Processing: Explorations in the Microstructure of Cognition, Vol. 2. McClelland, James L. and Rumelhart, David E. (eds.). Cambridge. MA: MIT Press: 216-271.

Winograd, Terry (1975) "Frame representations and the declarative/procedural controversy." In Respresentation and Understanding. Bobrow, Daniel G. and Collins, Allan (eds.). New York: Academic Press: 185-210.

Michael Masson is professor of psychology and obtained his Ph.D. in experimental psychology at the University of Colorado in 1979. He held a postdoctoral fellowship at Carnegie Mellon University and has been at the University of Victoria since 1980. During the 1986-87 academic year he was a visiting scientist in the human interface laboratory of Microelectronics and Computer Technology Corporation in Austin, Texas. His primary research interests are in memory and language comprehension and he also has published work on a connectionist model of context effects in word identification. He serves as a consulting editor for the Journal of Experimental Psychology: Learning, Memory, and Cognition, reviewed books on cognitive science and on human-computer interaction, co-authored a textbook on statistical analysis in behavioral science, and currently is co-editing a book on interdisciplinary approaches to the study of implicit memory.


Go to Semiotic Review of Books Home Page
Go to SRB Highlights
Go to SRB Archives