This excerpt from Connectionism and the Philosophy of Psychology. Terence E. Horgan and John Tienson. © 1996 The MIT Press. is provided in screen-viewable form for personal use only by members of MIT CogNet. Unauthorized use or dissemination of this information is expressly forbidden. If you have any questions about this material, please contact [email protected].

Chapter

1

lntroductio

~n andOverview

Connectionism arose very rapidly in the 1980sas a rival to the standard computer based approach to cognitive science. Connectionism quickly attracted a great deal of attention among philosophers , largely because it suggests the possibility of an alternative to the conception of mind on the model of the modern digital computer {the so- called computer metaphor ) a conception that had become prominent in thinking about the mind , both popular and philosophical , with the advent of the field of artificial intelligence in the 1950s. By reflecting on connectionism , and on the reasons it rose to prominence when it did , we have been led to a view of cognition and cognitive processing that is significantly different from familiar views of cognition , and also from the views of other philosophers who have reflected on connectionism . We believe that human (and other natural ) cognition is too rich and flexible to be simulated by computer programs . Hence, the computer model of the mind must be abandoned . However , this very richness of cognition requires mental representations that are syntactically structured . That is, cognition requires alanguage of thought , but one with processes that are not programmable or computable . We suggest that cognitive processes are effected by interacting cognitive forces that are systematically related to content. As aresult , there are many generalizations that are true of cognitive process es. But they are ceteris paribus generalizations from which the ceteris interaction paribus clause is ineliminable . Finally , the organization and the mathematical of cognitive forces are best understood in terms of the theory of dynamical systems. In this chapter we provide a sketch of this view and give an overview of the course of the argument we will pursue in the rest of the book . We begin by sketching the circumstances in which we arrived at the view . The picture of mind as computer dominated theory and methodology in artificial intelligence and cognitive science for a quarter of a century , and has deservedly been called the classical view in cognitive

2

Chapter 1

science. The most salient fact about computers is that they solve problems by virtue of being programmed with rules that refer to the problem domain . What the computer metaphor tells us, then , is that cognition is what the classical computer does: rule- governed symbol manipulation . Since this is held to apply to all cognition , it involves the claim that the human mind can in principle be perfectly imitated (indeed , duplicated ) by a computer program , with data structures corresponding to mental states such as beliefs and desires. One reason for philosophers ' long -standing interest in classical cognitive science is that it appeared to be the only way to approach the mind / body problem that would satisfy the requirements of the late twentieth century . The question of the relationship between the mind and the body has been one of the most difficult and widely discussed problems in philosophy since Descartes. As it appears today, the problem is one of finding a place for the mental in the natural world - of understanding ourselves both as beings that are part of the natural order studied by physics, chemistry, and biology and as intelligent agents whose actions can be understood only in the manner of the humanities and the human sciences. If , as scientific evidence suggests, the movements of a human body are always in principle causally explainable in terms of biochemical processes in the body and in the brain , how , then, can mentality make a difference? How can it be that human behavior is also explainable in terms of human goals, beliefs, plans, and purposes? What sorts of entities are human thoughts , feelings , beliefs, and so forth ? How , if at all , can such events or states take place in a physicochemical organ such as a human brain ? The computer metaphor appears to offer a way to understand how minds could be physically embodied and how mentality could matter to human behavior. The internal goings on of a computer can be correctly described and explained both at the level of electronic circuitry and at the level of executing a program . It was thought that the goings on in a human head could be correctly described and explained in directly' analogous ways - that the human mind could be construed as nature s computer program running on neural circuitry . But trouble arose in classical cognitive science. The emergence of connectionism in the 1980swas due in part to certain recalcitrant problems that had resisted treatment by the standard computer -based approach for more than a decade. These problems had, it is fair to say, generated a Kuhnian crisis in classical cognitive science. Connectionist architecture had been around for a long time , but in the 1970s it had received attention from only a few researchers.! In the 1980s researchers returned to neural networks in response to a felt need for an alternative approach .

Introduction and Overview

3

One reason for the philosophical interest in connectionism is that it way of understanding how suggests the possibility of an alternative 2 . Connectionist 'systems, or embodied be could physically mentality neural networks, present quite a different picture of the mind . A neural network consists of a large number of very simple neuron -like proces sors, called nodesor units, that send simple excitatory and inhibitory turned off by signals to one another. Nodes are turned on (activated) or nodes and connected of the on which depend" incoming signals, "output on the strength of connection ( weight ) between them. In practice, a or from given node may get input from just two or three other nodes3 several dozen. In principle , it could be thousands or millions . The most striking difference between such networks and conventional computers is the lack of a central executive. In a conventional the central computer the behavior of the whole system is control led at . In connectionist a stored unit systems there is program by processing no such executive and no program to control the operation of the system . All connections are local , so each node has only the information it information gets from the nodes to which it is connected. And the only that a node gets consists of simple input signals determined by the outputs of other nodes and by the weights between the nodes. No component of the system has more information than that . So, in particular , no " " whole is doing . part of the system knows what the system as a in the one at on what Furthermore , system is independent place goes of each node is behavior . The in the elsewhere on what of system goes determined only by its current state and its input . Nevertheless, connectionist systems can be interpreted as having of sentences interesting representational content analogous to the content nodes. over activation of in noun of or many , patterns phrases in the as understood be , can these weighted And retaining systems " " connections between nodes, the knowledge represented by representations that are not currently activated . Connectionist systems have been shown to be capable of solving interesting problems in some areas in which classical cognitive science has made little" progress, such as " , and to be capable of learning to solve problems pattern recognition " " in response to experience. A standard part of connectionist method " " are repeatweights ology is back propagation , a method by which " " until the edly adjusted in the direction of the correct, target output . the correct network produces output Connectionism was put forward in the 1980s as an alternative to classical cognitive science- that is, as offering an alternative conception of cognition . However , it is not clear that anything has emerged that could be called a, let alone the, connectionist conception of cognition . Classical cognitive science says that cognition is rule- governed

4

Chapter 1

symbol manipulation . As yet , connectionism has nothing to offer in place of the general conception this slogan represents. Two things were clear about connectionism from the outset. First , it was a vastly different way of doing things . One typically does not program a connectionist system; instead one gives it a training regimen , from which it learns to solve problems . As a result , second, connectionism suggested quite different ways of thinking about particular cognitive processes and abilities , such as perceptual recognition or language learning . But this does not necessarily result in a conception of cognition . Connectionism is praised , for example, for being an architecture that is more neurally plausible , or more brain -like . The question is: What does the fact that a model is like a brain tell us about cognition ? The mystery is how any physical object can have cognitive states. Connectionism says that thinking is activity in a neural network . But surely not all of the activity in any neural network would count as . mental in any sense. A network could be too small to do anything remotely resembling human cognition ; a large network could have activity so uninteresting as to count as, at best, brain dead. A connectionist conception of cognition would have to tell us what it is about network activity that makes it mental activity , what kind of network activity constitutes mental activity , and why that kind of activity constitutes mental activity . There were no clear answers to such questions in the early connectionist rhetoric of difference. In part for this reason, connectionism attracted philosophical fans of such different philosophical persuasions as Paul and Patricia Churchland , on one hand , and Hubert and Stuart Dreyfus , on the other. Most early philosophical discussion of connectionism was concerned in one way or another with the question of whether it could , replace in whole or in part , the classical, computational conception of mind . This involved foundational questions: What are, or what should be, connectionism ' s fundamental assumptions about the nature of mind , about the nature and structure of mental states, and about the character of mental processes? How would such assumptions differ from those of classical cognitive science? How plausible is it to give up classicism' s fundamental assumptions , in view of the lack of success, in the history of psychology, of earlier approaches such as behaviorism and associationism, to which connectionism is frequently compared ? Philosophers have, for the most part , approached these questions by examining the systems, methods , and formalisms that have been " " developed by connectionists and asking What have we here? There is, of course, a need for careful examination of connectionism as it exists today. We have, however , approached matters somewhat differ -

Introduction and Overview

5

as we find ently in this book . Instead of beginning with connectionism classical the of , computer -based it , we have asked what features in view of those approach led to the problems it experienced, and why , looked in promising (aside from its problems , connectionism particular we have asked what general . Thus , being the only other game in town ) treatment in the classical successful eluded features of cognition have we have asked whether And so. done have paradigm , and why they will fare better on connectionism that there are reasons for believing will have to be connectionism how and those features of cognition , better. developed if it is to fare One feature of cognition on which connectionists claim an advantage is multiple simultaneous soft constraint satisfaction . Suppose you are shopping for a car. You will , no doubt , consider performance , style , more. (Each of safety, price , quality , availability of service, and perhaps course.) of factors different of , Intuitively , these nouns covers a number to buy . In which car in you consider all these factors together deciding be can considerations the all put into the a connectionist system, too , work . their do to allowed and to cognitive hopper at once, so speak, constrains automobile in an for are Each of the features you looking can be your final decision . But most of these constraints are soft they its is job properly . violated when your decision making system doing is, if price is that can much how If there is a strict bound on pay you that all to be able not will you desire a hard constraint then you get combination best the You so on. and in safety and style and performance , get . without do must some but can that , of desired features you you be constraint be a soft ; you might On the other hand , price itself might able to go somewhat over your budget to get certain features you want . Thus , the same thing - in this case, a price limit - can be a hard constraint in one context and a soft constraint in another. Classical cognitive science need not entirely ignore the softness of it must attempt to obtain multiple constraints . However , it must fake it ; from satisfaction of multiple soft constraints processing describable by is no real cognitive softness there hard rules of cognition . Upon analysis, science in the system. Thus, classical cognitive postulates that the . hardness from apparent softness of cognition emerges to be seems Human cognition , however , naturally good at accommodating but also their softness constraints of not only the multiplicity in kind from different is . This suggests that human intelligence " " mode of satisfying our that it In particular , suggests computer intelligence . . derived than rather basic is multiple soft constraints comes satisfaction constraint naturally to Multiple simultaneous soft the into are constraints the cognitive connectionist models. When all put of the models . hopper, some may get overruled in the solution Many

Chapter 1

described in the ParallelDistributed Processingvolumes involve multiple simultaneous soft constraint satisfaction. Thus, multiple simultaneous soft constraints pose problems for classicism , and connectionism seemsnaturally well suited for dealing with this and other " soft" aspects of cognition .4 In this book we emphasize the softness of cognition in general, seeing soft constraint satisfaction as one instance of a ubiquitous feature of cognition . The generalizations that can be made about cognitive processes have exceptions without limit within the realm of psychology .5 These exceptions depend upon content, not form . Hence, all generalizations about cognitive processes are inevitably soft; generalizations about cognitive processes are not formalizable by means of tractable algorithms . The observed limitations and failures to scale up within classical cognitive science foretell in -principle limitations to the approach . In chapter 3 we point to certain ways in which human cognition . appears to be open ended. It is open ended with respect to the kinds of things it can represent, and virtually any pair of things it can represent might be recognized to be relevant to each other under some circumstance . This is why generalizations in psychology cannot be made exceptionless . There are also certain human capacities that are open ended, most notably the ability to use natural language appropriately . Finally, human cognition is open ended with respect to the kinds of activities it can engage in intelligently . There is, for example, no limit to the kinds of games within which a human being can employ reasonablestrategy. There is no known or foreseeableway to provide for these kinds of open-endednessby means of programmable rules. Indeed, these kinds of open-endedness do not appear to be amenable to generation by recursive rules, which is the kind of open-endednessthat comes naturally to the rules-and-representations paradigm . Our conclusion is that it is very unlikely that human cognitive capacities are achievable by rules of the kind postulated by classicism- programmable rules that refer to (representations of) items in the domains of the cognitive tasks the programs deal with .6 And by this we mean that human cognition cannot be simulated by programmable rules; the distinction between hard wiring and explicitly represented rules is not relevant to this point . Furthermore , certain cognitive processes depend upon holistic properties of belief systems, such as relative simplicity and conservatism. Indeed, relevance itself is a holistic property . There is no known formalism for taking such holistic matters into account for belief systems as large as human belief systems. There are, therefore, adequate grounds for attempting to develop an alternative conception of cognition . Human cognition is too rich and too flexible to be captured by hard rules. But becauseof this very richness, human cognition (and that of the

Introduction and Overview

7

higher nonhuman animals) requires complex representations that are systematically constructed from constituents - that is, cognition requires syntax, a language of thought . But this does not necessarilymean syntax as it is conceived in classical cognitive science or in formal logic . Constituents need not be parts of complex representations. What is necessary is that all complex representations with a constituent of a certain kind be related to that constituent in the same way, so that given the constituents the complex representation is determined automatically . " Thus, at one time we called our view Representations without " Rules, meaning syntactically structured representations but no exceptionless rules adverting to those representations. This does not make cognition chaotic or unruly . There are many generalizations that are true of cognitive systems and kinds of cognitive systems. But these generalizations are inevitably ceterisparibus generalizations. There are always exceptions to any generalization that is of sufficient scope to be even moderately interesting . And , because of the open-endednessof cognition and the (content- dependent) potential relevance of anything to anything , these exceptions cannot be fully spelled out to replace the ceteris paribusgeneralization by a universal generalization of classical logic . However , we came to see that merely elaborating the picture as sketched so far is unsatisfactory . On the classical view , it is possible to understand how cognitive processes can be sensitive to the structured content of representations: cognitive processes are effected by automated rules . Thus , programmable representation -level (PRU rules have a form that could give insight into how cognitive pr ~ essesoccur. Ceterisparibus generalizations , on the other hand , merely record what happens (and what , counterfactually , would happen), for the most part . In themselves they give no insight into how things happen . Thus, we need a story about cognitive processes. Connectionist cognitive processes are effected by spreading activation . Forces are emitted by representations. These forces tend to activate or inhibit other representations. We take the idea of cognitiveforces seriously. In connectionism , cognitive forces are subserved by forces emitted by particular nodes. With distributed representations, however , the relationship between cognitive forces and the forces emitted by the nodes of activated representations can be quite abstract. Cognitive forces cannot be identified with node-level forces. Looked at from the cognitive point of view , the most basic and central cognitive processes are effected by content appropriate cognitive forces. These forces are often determined by relations among representations which conspire to produce a force. A desire for a cup of coffee can conspire with a belief that there is hot coffee in the next room to produce a force that tends to move the cognitive system to a certain

8

Chapter1

output state. Beliefs conspire with other beliefs, with desires, with fears, and so forth systematically over vast ranges of possible beliefs and emotions . Hence, there are systematic generalizations concerning the (cognitive ) directions in which cognitive systems tend to move. However , cognitive forces can come into conflict . You may not want to talk to a person who is in the next room . This may produce a force strong enough to cause you to forgo that cup of coffee. Thus, cognitive states, and conspiring combinations of cognitive states, produce forces that result in causal tendencies within cognitive systems. But they can be defeasibletendencies only . Any such causal tendency could be overridden by a stronger force or combination of forces. Hence, the systematic generalizations about cognitive processes will have exceptions. We call these systematic ceteris paribus generalizations " soft laws." Finally , where cognitive forces can conflict , they can also cooperate, so that forces from two distinct cognitive states tend to push the cognitive system in the same direction . We believe that the mathematical theory of dynamical systemsis a useful framework for understanding the organization and the interaction of cognitive forces. Dynamical Systems theory is also the natural mathematical framework for characterizing neural networks. Thus, the mathematical theory of dynamical systems provides a needed link between a neural network and the cognitive system it subserves, or, in terms of connectionism, between a connectionist network and the cognitive processes or capacitiesthat the network is said to model. Dynamical Systemstheory provides a way of understanding how a network can subservea cognitive system and how cognitive forces can be subserved by a physical network . Happily , many connectionist ideas fit naturally into this picture . To summarize the picture : (1) Natural cognitive systems need a rich structure of available representations. Thus, representations must be constructed as needed, in accordancewith some kind of syntax (but not necessarily classical, part / whole syntax ). (2) Cognition is too rich to be simulated by representation -level rules. However , (3) there are many ceteris paribus generalizations that apply to cognitive systems. (4) Cognitive processes are effected by node-level cognitive forces, which are subserved by but not identical with forces in a physically realizable network . Cognitive forces often result from conspiring cognitive states, and can compete and cooperate. Finally , (5) dynamical systems provide the crucial mathematical / design link between cognitive systems and the networks that realize them. In chapter 2 we use an influential characterization of classical cognitive science to bring out the fundamental assumptions of classicism. This characterization , due to David Marr (1982), identifies three interrelated

Introduction

and Overview

9

levels of description that figure importantly in classical cognitive science : (1) the level of the cognitive qua cognitive , at which cognitive transition functions are specified ; (2) a middle level , which specifies algorithms by which cognitive -transition functions are computed ; and (3) the level of physical implementation of the algorithm . The most important assumptions of classicism are (1) that representations have structure (typically syntactic ), (2) that cognitive processes conform to programmable representation level rules , and therefore, (3) that cognitive processes conform to tractably computable transition functions . In chapter 3 we argue that classical cognitive science presents a fundamentally mistaken vision of human cognition . We point to various ways in which human cognition is open ended and argue that these kinds of open-endedness are not attainable by means of programmable rules. The most basic way in which human cognition is open ended is in its capacity to represent. This capacity has led to problems for classical cognitive science(including the so- called frame problem ) that pertain to relevance : updating memory in response to changing information , finding information in memory that is relevant to the current task, performing tasks that require information from several domains , and so on. We argue that such problems are not solvable in the classical paradigm because of the interaction of two features of human cognition : (1) the fact that there is no limit to the number of different things , states, properties , events, etc. that a human being can represent and (2) the fact that anything a human being can represent could , in some circumstances, be relevant to any other thing that human can represent- the potential relevance of anything to anything . It is important that relevance is, in general, a matter of contentof representations , typically in ways not directly mirrored in their form . These considerations strongly suggest, we argue, that it is not possible to codify - in the precise and exceptionless way required for a program - all of the possible exceptions to any generalization concerning human cognition . If this is correct, human cognition cannot be duplicated by a programmable computer . It also means that programmable computers should not be expected to be good at many of the tasks most characteristic of human cognition , such as perceptual recognition and categorization under vague and multifaceted concepts. We do not claim that this argument is conclusive. It is a broadly empirical argument . But we think it is a strong empirical argument . In any case, in view of the crisis in classical cognitive science, it is appropriate to attempt to articulate alternative conceptions of cognition . Thus, we take the negative conclusions of chapter 3 for granted and proceed on the assumption that classical cognitive science does not provide a correct conception of human cognition .

10

Chapter 1

In rejecting classicism as a conception of human cognition , we do not mean to deny the value of classical cognitive scienceas an ongoing research program . A great deal can be and has been learned about cognitive problems by writing programs aimed at solving them , and a great deal is revealed about theories (and their defects) by implementing them classically. It is quite likely that these things cannot be accomplished in other ways . As far as we know , no one has learned anything comparable about any cognitive or philosophical problem from connectionism . The effort to solve cognitive problems by classical methods may well remain forever an unequalled route to understanding the nature of particular cognitive problems . In chapter 4 we sketch an alternative framework for cognition : Noncomputable Dynamical Cognition , in which dynamical systems replace algorithms at the middle , mathematical / design level . We begin the chapter by generalizing Marr ' s three- level framework for classical cognitive science, which yields (1) cognitive -state transitions (but not necessarily tractably computable cognitive -transition functions ), (2) mathematical -state transitions (but not necessarily algorithms ), and (3) implementation . We then set forth a succession of approaches to mentality that conform to this general three- level framework while deviating increasingly from classicism ; Noncomputable Dynamical Cognition is the most radical of these nonclassical alternatives . Chapters 5- 8 are concerned primarily with the top (cognitive ) level. Chapter 9 then ties this discussion to the middle (dynamical systems) level . Classical cognitive science is a package deal, involving syntactically structured representations and rules referring to those representations. In chapter 5 we argue that cognitive systems must have syntactically structured representations that make a difference in cognitive processing - what we call effectivesyntax . Thus, what is to be rejected from the classical package is rules that refer to those representations (PRL rules ). It is necessary to argue anew for syntactic structure because many connectionists, and many philosophers who have rejected classicism, have wanted to throw out syntax with classical cognitive science, and becausemany arguments for syntactic structure have a distinctly classical flavor . Our main argument for syntactic structure , which we call the tracking argument , concerns getting around in the world . We discuss the capacity to represent sufficiently any immediate locale one might be in , and the cognitive demands of complex physical skills . To get around in the world , a cognizer must keeptrack of enduring individuals that have changing , repeatable properties and relations. Doing this requires that mental predicates be applied to mental subjects, and it requires the

Introduction and Overview

11

capacity to apply predicates to subjects on a vast scale. Some philosophers may seehere only an argument for compositional semantics, not an argument for syntax . But , we argue, there is no way for complex , natural cognizers to have representations with compositional semantics on the vast scale they do have except by having complex representations that are systematically related to simpler representations - i .e., by having representations with syntactic structure . Thus, we argue that there must be a languageof thought , although this may be (and within connectionism almost certainly is) a nonclassical language in which the relation of constituent to complex representation is not part to whole . We describe two methods for encoding structure nonclassically in connectionist networks : recursive auto- associative memory (in chapter 4) and tensor products (in chapter 5). We also discuss some progress that has been made in basing structure-sensitive processing on such representations . In chapter 5, deductive reasoning provides a secondary argument for syntax : natural cognizers engage in deductive reasoning (and in formal fallacies), and they could not systematically reason in accordance with formal principles if their representational states did not have suitable form (syntax ).7 In chapter 6 we introduce cognitive forces, conspiracy among cognitive states to generate cognitive forces, and cooperation and competition among cognitive forces. We lead up to this by arguing that causal tendencies in cognition are defeasible, and that possible defeaters cannot be completely specified in terms of form or content. In any case, even if all possible defeaters could be spelled out , defeasible causal tendencies are properly expressed by ceterisparibus generalizations . Chapter 7 contrasts soft laws with the standard conception of scientific laws , according to which a law must be exceptionless, or, in the special sciences, at least quasi exceptionless (that is, having no exceptions within its own domain ). We also point out in chapter 7 that the laws comes not merely from universal quantification generality of scientific "" from but , parameters that is, quantification over (infinite ) arrays of determinate properties of a determinable type (e.g ., length or mass). This applies both to the laws of the physical sciencesand to the soft laws of cognition , which have intentional rather than quantitative parameters. We conclude chapter 7 with an explanation of why the laws of psychology are different from the laws of physics . The main business of chapter 8 is to argue that ceterisparibus generalizations , including soft laws , are not empirically disreputable , as many have thought them to be. Ceterisparibus generalizations are subject to empirical confirmation and disconfirmation , and they can figure in explanations in the ways expected of scientific laws - particularly in

12

Chapter1

deductive / nomological explanations and functional explanations . In that chapter we also point out some implications of the picture of cognition that we have been sketching for psychological theory and philo sophical methodology . In chapter 9 we elaborate the Dynamical Cognition (DC ) framework for cognition that was proposed in chapter 4, incorporating themes from chapters 5- 8 and also incorporating certain suggestive ideas about cognition that have emerged from connectionism . We explain how the idea of cognitive forces not only fits easily into the DC framework but also receives a natural explication in that framework . We also discuss how some classicist design principles centering around effective syntax might be incorporated into the DC framework . And we consider the prospects for dealing , within the DC framework , with the problems for classicism discussed in chapter 3 and elsewhere in the book . In some cases, the DC framework provides suggestive ways of thinking about the problem . In other cases, we have nothing new to " suggest. In the instances where DC is suggestive, we do not say Here " " is how it works ; we say Here is a picture of cognition in terms of which to think about the problem , and a way to think about it in terms of that picture ." As Jerry Fodor (1983, p . 107) said in a similar context (specifically about analogical reasoning), " nobody knows anything about how it works ; not even in the dim , in -a-glass-darkly sort of way ." About some of the aspects of cognition that we discuss below, even dim , in -a-glass-darkly ideas are progress, especially in the context of a systematic conception of cognition . Though we frequently focus on connectionism for concreteness, our concern is not to defend connectionism per Sf, but to sketch and begin to elaborate a new view of cognition - a view that is, in part , informed by connectionism . Human -like cognition might only be subservable by physical systems quite different in nature from current connectionist networks ; in principle , our discussion would carry over, mutatis mutandis, to such alternative systems. We conclude that , if connectionism is to yield a viable alternative conception of cognition , three things are necessary: (1) It needs syntactic structure , and processing richly sensitive to that structure , but (2) this processing must not be describable by PRL rules. (3) It must make progress on the problems that appear to be beyond the capabilities of classical cognitive science. It has promising features that suggest that it may be able to deal with those problems and exhibit the requisite softness if it can support effective syntax . There are at least some ideas and some minimal models in existence that suggest that it may be able to support effective syntax .

Introduction and Overview

13

In this introductory chapter and throughout the book , we have taken representation for granted . In doing so, we simply follow connectionist (and classicist) practice . Representations are as much a part of connectionist theory as are activation levels, weights , and update equations . Every connectionist model has representations; otherwise it would not be a cognitive model .8 In both classical and connectionist , content is assigned to certain states, essentially by the mod modeling ' eler s fiat - with the modeler aiming , of course, to respect the causal properties of the system. Models are described in terms of ' representations , and cognitive processing is construed as the system s evolution from one representational state to another. It is assumed that natural cognitive systems have intentional states with content that is not derived from the interpreting activity of other intentional agents. But it is not the business of either classical cognitive science or connectionist " " cognitive science to say where underived intentionality comes from (although each may place certain constraints on an answer to this question ). Thus, intentionality as such is not at issue in this book , and it will be discussed only as it comes up in relation to other issues.

This excerpt from Connectionism and the Philosophy of Psychology. Terence E. Horgan and John Tienson. © 1996 The MIT Press. is provided in screen-viewable form for personal use only by members of MIT CogNet. Unauthorized use or dissemination of this information is expressly forbidden. If you have any questions about this material, please contact [email protected].