Dealing in Futures: Folk Psychology and the Role of Representations in Cognitive Science. CSRP 289 Andy Clark School of Cognitive and Computing Science ÑÇÖÞÇéÉ« Brighton, BN1 9QH andycl@cogs.sussex.ac.uk ABSTRACT. Two famous theses distinguish the Churchlands' longstanding speculations on the scientific futures market. The theses are (i) that there is something rotten at the heart of folk psychology, and (ii) that the march of computational and neuroscientific investigation will lead us to embrace a set of distinctly different explanatory and representational resources for understanding the mind. I argue in favour of the latter proposition, but against its use as an argument against folk psychology. More strongly, I suggest that it is against the backdrop of an increasingly alien and fragmentary inner scientific story that the virtues of the folk framework become most apparent. The scientific advances upon which the Churchlands so ably draw will have their most profound impact not upon our assessment of the folk discourse but upon our conception of the role of representations in the explanatory projects of cognitive science. Representation, I suggest, will indeed be reconceived, somewhat marginalized, and will emerge as at best one of the objects of cognitive scientific explanation rather than as its foundation. to appear in R.N.McCauley (Ed.) THE CHURCHLANDS AND THEIR CRITICS. 0. Folk Psychology loses, Neuroscience gains? For well over a decade now, P.S. and P.M. Churchland have been prominent speculators on the scientific futures market. Two of their most important targets can be identified as (1) the attempt to establish the low long term value of the framework of daily mentalistic explanations now known (a little unfortunately) as folk psychology, and (2) the attempt to demonstrate that the scientific study of cognition will increasingly deal in the somewhat different representational and explanatory resources emerging from work in neuroscience and broadly connectionist kinds of computational modelling. Though both the Churchlands are sympathetic to each of the two projects, it is probably fair to say that it is P.M.Churchland who has made much of the running on (1) (see especially his 1979, 1981, 1989 ) and P.S.Churchland who has pursued the most detailed investigations of (2) (see her 1986 and 1992 , the latter being co-authored with the neuroscientist T. Sejnowski). The separation is nonetheless, clearly artificial - a full half of P.M.Churchland's 1989 is devoted to matters which concern the explanatory resources of connectionism and neuroscience while works such as P.S.Churchland 1978, 1980, and 1986 leave us in no doubt that she too shares a deep-rooted pessimism about the ultimate integrity of folk psychology. Intense philosophical interest in the claims about folk psychology has, I believe, served to obscure some of the Churchlands important observations concerning the changing nature of the explanatory resources of the scientific study of mind. This is unfortunate because, in my view at least, the two issues are more or less orthogonal and the insights concerning the methodology and explanatory profile of future Cognitive Science should be valued even by those (myself included) who reject their strong claims concerning the poverty of the folk framework. My strategy will therefore be to pursue each of these issues in its own right. I begin (section 1) by briefly summarizing some of the results and conjectures which play a key role in both debates. These concern the nature of representation and learning in recent, neurally inspired, models of mental processing.The potential impact of such results on both Cognitive scientific practice and our assessment of folk psychology are briefly discussed. Section 2 homes in on issues concerning the future role of representation-talk in Cognitive Science. In this section I attempt to add even more fuel to the Churchlands' fire by highlighting the availability of a new and stronger model of learning and conceptual change, and introducing some recent studies concerning the effects of the interplay of evolution and individual learning. The emerging lesson - which I see as a central implication of the Churchlands' overall view - is that we may well need to re-think the role of representations in the future explanatory projects of Cognitive Science. Representations will be frequently alien (unrelated to folk items), sometimes inward-looking (concerning states of other neural systems), module-specific (no global or 'central' cognitive code) and as much the objects of our explanatory endeavours as their foundation. I end (section 3) by returning to the issue concerning the probable lack of fit between the ontology and explanatory resources of a future cognitive science and those of folk psychology. This lack of fit is, I argue, paradoxically- our best clue to the peculiar virtues of the folk framework. In particular, I suggest (building on some brief comments made in section 1) that the folk are best seen as engaged in a kind of global knowledge-detection and that scepticism about the value of the folk enterprise therefore cannot be justified by appeal to results which rather concern the form of representation in cognitive SUB-SYSTEMS. More positively, I suggest that once we are clear that the mere lack of inner quasi-linguistic items is not itself sufficient to undermine the folk discourse, the way is open to build a balanced image of the folk project. Once such an image is in place it becomes clear that the most likely impact of scientific advance takes the form not of an Eliminative but of an Augmentative materialism. 1. The Impact of Non-Sententialism. The Churchlands' philosophical projects are powerfully rooted in ongoing work in neuroscience and in (broadly) connectionist Artificial Intelligence. Two aspects of this recent work bear centrally on the topics of this paper. The first directly concerns the nature of the story about internal representational states which is emerging from such studies: it is a story which challenges any deeply sentential or quasi-linguistic (symbol system and grammar) vision of the inner cognitive economy. The second (related) aspect concerns the nature and scope of processes of individual learning. The new representational forms which challenge the sentential vision can emerge as the products of learning processes which constitute existence proofs of the broad possibility of types of learning which outstrip the kinds of learning associated with the classical Fodorian vision of learning i.e learning as a process of hypothesis generation and testing using the restricted (and restricting) resources of a fixed innate symbol system (see Fodor (1975)). A word on each. Regarding the basic form of internal representation, both P.S. and P.M. Churchland stress that the inner representational resources posited by Connectionist Artificial Intelligence do not look much like quasi-linguistic declarative structures. Seen through the lens of connectionist A.I. "the basic kinematics of cognitive creatures is a kinematics not of sentences but of high-dimensional activation vectors being transformed into other such vectors by passing through large arrays of synaptic connections" ( P.M. Churchland, 1989, p.xvi). In similar vein, P.S. Churchland, writing with the neuroscientist T.Sejnowski, comments that: Instead of starting from the old sentence-logic model, we model information processing in terms of the trajectory of a complex non-linear dynamical system in a very high-dimensional space" (P.S.Churchland and T.Sejnowski, 1990, p.234). This is not the place to review the detailed and ever-growing body of evidence for such claims. Suffice to say that it is widely accepted, even among its detractors, that connectionist A.I. really does offer a new and in some genuine sense non-sentential vision of inner processing (for fuller descriptions see P.M. Churchland (1989), Clark (1989)). But suppose then that some such non-sentential model really does characterize information processing in the brain. What would follow, either for Cognitive Science or for the folk vision of the mind? The implications for Cognitive Science are clearly direct and profound. Called into question would be the attempt to model reasoning as a kind of proof theory defined over internal data structures (see Oaksford and Chater (1991)) and with it the very notion of a clean boundary separating perceptual and cognitive processes (see P.M. Churchland (1989) chapter 10). At the same time the usefulness of familiar distinctions (processing/data, algorithm/ implementation) would be challenged (see P.S. Churchland and T.J.Sejnowski (1990) and Clark (1990)). Even the once-unimpeachable methodology of seeking a symbolic competence theory as both the starting point and explanatory core of our Cognitive Scientific endeavours is revealed as embodying illegitimate assumptions (see Clark (1990)). Serious implications indeed. By contrast, the implications for the folk framework (or more properly, for our considered assessment of the integrity of that framework) are far less obvious. The folk do indeed use sentential formulations as a means of specifying each others mental states. Thus I may say of Pepa that she believes that the sun will shine. And in so doing I rely on a proposition embedded in a that-clause and expressed by a sentence of English. But what can this reasonably be supposed to commit me to (philosophically) as regards the form of internal representation in Pepa's head? In general it is certainly NOT the case that the use of a sentential description language commits one to the existence of even quasi-sentential structure IN the domain described (think of sentences about protein folding, or about geology, or about cooking,..). On the face of it, at least, the very MOST that the sentential folk discourse might seem to saddle us with is the commitment to there being some kind of internal state or states which are causally potent and which somehow support the specific mental contents allude to in a given (supposedly true) folk psychological account. (And see Van Gelder (1993) and Clark (forthcoming-a) for some doubts broadly Rylean doubts (Ryle (1949)) even about this!) But this latter kind of claim is not obviously (footnote 1) contradicted by work in the connectionist or neuroscientific paradigms. Thus we might reflect that many existing connectionist networks happily succumb to useful knowledge ascriptions couched in the folk vocabulary. Thus P.M.Churchland (1989, ch.9) describes a single net which is relatively happily described, at times, as believing that an object is a mine and not a rock. By this I mean that, although the system is surely too impoverished to count as REALLY believing that it is confronting a mine (op.cit. p.177), our doubts about the belief-invoking description are not rooted in its non-sentential character: we would have equal reservations about a classical, rule and symbol device which could (only) partition inputs into mine and rock indicating signals. Similarly, P.M. Churchland is seen to claim that a good connectionist encoding might capture the prototypical structure of fine-grained types of situation whose contents he is willing to describe as e.g. depicting a mouse eating sesame seeds or hickory nuts (P.M.Churchland, 1989. ch.10, p.207). It is thus hard (or so I claim) to see what can be amiss, in such cases, with the folk psychological framework - the creature whose overall behaviours are as complex as our own (and this MUST be on the cards if connectionism is put forward as a potential new model of human thought) and whose inner connectionist network includes a distinctive subvolume devoted to representing the fact that the mouse is eating sesame seeds is surely a being ripe for folk psychological description. DOES P.M Churchland depict the mere non-sentential form of connectionist internal representation as inimical, in and of itself, to the integrity of the folk framework? I am not sure. In a revealing passage originally published as part of a commentary on D. Dennett's book THE INTENTIONAL STANCE he writes that: We both (i.e. D.Dennett and P.M.Churchland) accept the premise that neuroscience is unlikely to find "sentences in the head" or anything else that answers to the structure ofindividual beliefs and desires. On the strength of this shared assumption I am willing to infer that folk psychology is false and that its ontology is chimerical. P.M.Churchland (1989) p.125 It seems, from this passage at least, that the MERE failure of a putative internal representational system to evince some kind of sentential form is not itself sufficient to compromise folk psychology. Instead we would need to discover that nothing in the scientific description "answers to the structure of individual beliefs and desires". But if we are faced with a connectionist system in which a distinctive subvolume of activation space is dedicated to representing e.g. that the mouse is eating sesame seeds it is unclear in what sense we lack an inner item which 'answers to' the description of the being as eg believing that the mouse is eating sesame seeds. Perhaps, however, we should be taking these descriptions (of subvolumes of activation space as representing such familiar states of affairs) with a pinch of salt. My own view (developed at length in Clark (forthcoming-a)) is that it is unrealistic to expect subvolumes in the activation space of single networks to turn out to encode such familiar contents. Instead, the contents which figure in the folk discourse may refer, at best, to rather global properties of systems (like us) comprising multiple networks whose individual tasks are to represent specific aspects of the input: aspects which are not postulated as components of the folk- individuated mental states. Some partial and preliminary evidence for such a view is already available in the range of unexpected deficits and dissociations revealed by studies of brain damaged patients. Thus the Churchlands observe elsewhere that "the accidental destruction of isolated brain areas (lesions) leaves people with isolated and often very curious cognitive deficits" (P.M. and P.S.Churchland (1990) p.309). Indeed so - see e.g. Ellis and Young (1 988), Warrington and McCarthy (1990), Shallice (1988), Humphreys and Piddock (1987) and P.S.Churchland and T.Sejnowski (1992) ch.5. In the inner realm, folk psychological items (knowledge of persons, recognition of faces, identification of everyday objects) fragment. But even this, I claim, fails to undermine the folk ontology of mental states. Thus imagine now a sophisticated connectionist device whose range of behaviours (verbal, sensor-motor, recognitional etc.) is so great as to incline us to say that it really does know about rocks and mines i.e. it is not an impoverished single net but a complex organism comprising multiple input channels and processing modules. Such a device has, I would like to say, a variety of global skills: as a result it is able (courtesy of the operation within it of some multitude of networks which act as mini-experts dealing with unintuitive fragments of the gross tasks which confront the overall system) to satisfy us as to its right to be described as knowing e.g.that such and such an input signifies the presence of a mine. In such a case, any inner fractionability is surely neither here nor there: what the folk talk is a comment upon is rather some global complex of capacities and skills not even plausibly subserved by any inner resource. Potential fragmentation is thus the empirical price of the complex of behaviours we rightly DEMAND before engaging in the full-blooded use of the folk vocabulary. The inner economy can be both non-sentential and fragmentary/distributed relative to the folks descriptive practice without that practice being in any obvious way compromised. I conclude that we do not really have any clear sense of what it would be like to discover, scientifically, that nothing inside us "answers to the structure of individual beliefs and desires". We DO know what it would be like to discover that the inner story is non-sentential. And we DO know what it would be like to discover that our inner resources divide up the task of representing the world in ways that the folk practice would never lead us to predict. But once we see the folk practice as commenting on the overall body of skills and knowledge which our daily behaviour exhibits,even the latter discovery seems compatible with the truth of the folk descriptions. To say of someone that they believe that such and such is, on this account, to locate them in a web of possible actions and responses whose inner roots may be both non-sentential and highly various. The inner facts which "answer to the structure of individual beliefs and desires" are thus not LOCALIZED inner facts. They are not facts about some specific inner SUB-SYSTEM. One (misguided) source of apparent pressure on the folk framework is surely the drive to identify beliefs and desires (and all the rest) with scientifically respectable SUB-STATES of the overall cognitive system. This expectation is fuelled, I believe, by the ubiquitous talk, within Cognitive Science of the inner VEHICLES of folk psychological content. The folk explanatory framework, however, leaves it quite open that the only vehicles of THOSE kinds of content are whole situated organisms. This kind of view is clearly expressed in Ryle (1949) and echoed -with some unnecessary instrumentalist baggage- in Dennett (1978) (1987). It is also pursued, in somewhat different ways, in Van Gelder (1993) and in Clark (forthcoming). Time now to shift the focus to our second issue viz the nature and scope of learning. Although the issues concerning folk psychology can be raised again here, I do not propose to do so. P.M.Churchland does, famously, claim that the failure of folk psychology to itself illuminate processes of learning and conceptual change constitutes another reason to question its integrity (see P.M.Churchland (1981) reprinted in his (1989) pp.6-9). But it has been argued elsewhere (Clark (1989) ch.3, Horgan and Woodward (1990)) that these alleged explanatory failures are of uncertain import as it is entirely plausible to insist that the folk discourse was never committed to the explanation of such phenomena in the first place. I do not propose to cover this terrain again here. Instead, I propose to highlight an issue concerning the potential impact of alternative scientific models of learning on our conception of the explanatory project of Cognitive Science. The place to start is with the broadly Fodorian view that there really is not (and cannot be) any such process as radical but rational conceptual change and learning. Instead, all such 'change' always involves the recombination of originally innate representational resources. The problem of explaining the origins of these representations is then shunted onto biological species level evolution (see Fodor (1975)). Both Churchlands are (rightly) critical of this evasive tactic. It certainly LOOKS as if human beings engage in processes of real representational change and development, and passing the representation- generating buck entirely to evolution seems merely to postpone the problem rather than to solve it (see P.S.Churchland (1978)). The good news (and for lots more on this see Clark (forthcoming )) is that connectionist learning algorithms offer a clear existence proof (and this, of course, remains DESPITE the biological implausibility of the most popular algorithms) of the general possibility of representation-generating yet rational learning and hence of the possibility of representational change unbounded by the expressive limitations of a pre-existing symbol system. base. This is because such systems are provably capable of inducing new representations to facilitate success at a problem-solving task. In addition, even if some innate representational resources ARE provided, the form of learning does not limit the future representational growth of the system to the quasi-logical manipulation and recombination of those resources (see Rumelhart and McClelland (1986a) p.141; Clark (forthcoming-a) ch.2). Connectionist approaches to learning thus extend a fascinating invitation to try to understand the non-representational origins of representation itself. This project will, I predict, increasingly come to dominate Cognitive science. By contrast, most of the work of traditional cognitive science consisted in trying to discover/guess what the mature representational resources of the brain looked like. Once the representations were on the table, most of the work of classical, Fodorian cognitive science would have been finished (see e.g. Fodor (1987b) p.147). The new explanatory endeavour clearly embodies a new conception of the ASPIRATIONS of Cognitive Science insofar as it is at heart a representation-generating, and not a representation-assuming, approach. Such a framework provides an exciting opportunity to begin to probe forms of representational content which are very far removed both from those specified in familiar folk psychological talk and from those which have figured in previous competence- theoretic speculation. In the next section I try to capture some of the flavour of these new possibilities. 2. The Role of Representation in the Future of Cognitive Science. The scientific study of cognition is replete with talk of representation. What I find most suggestive in the Churchlands' work is the persistent attempt to challenge our conception both of representation itself and of its role in the explanatory projects of cognitive science. To illustrate this, I shall briefly rehearse one aspect of their treatment, and then offer an illustration which may help reveal the potential radicalness of the proposed reconception. In a co-authored paper (P.M. and P.S.Churchland (1981)) the Churchlands draw a useful contrast between what they term calibrational versus translational kinds of content. The ascription of translational content involves first the mapping of a whole system of representations onto the system embodied in a public human language and second the systematic translation, in the light of that mapping, of propositional contents between the two representational schemes. Calibrational content, by contrast, is assigned on a much simpler and less globally systematic basis viz. a calibrational content of X can be assigned to any physical state which repeatedly and reliably indicates the presence of the feature X in the systems accessible environment. Thus we assign the calibrational content of 'temperature O'c' to a certain height in a column of red alcohol. (See P.M. and P.S. Churchland (1981) reprinted in P.M.Churchland (1989) pp. 42-43.) Given the putative absence of any quasi-sentential inner code, it becomes plausible suppose that the basic method of assigning contents to the states of inner sub-systems will need to be calibrational rather than translational. One immediate benefit of giving up on the search for translational contents 'in the head' is that it opens the door to a far-reaching multiplicity of internal representational systems. We no longer need search for a single representational system able to underpin all processes of rational learning and higher cognition. Instead, it seems likely that we deploy an integrated hierarchy of quite different computational/representational systems (P.M. and P.S. Churchland (1990) p.309). Once we give up the dream of translational content, we can treat each such system in its own right as a target for the assignment of calibrational content. In addition, the contents of calibrational contents can be arbitrarily removed from those of daily talk, whereas translational content is required to connect directly and systematically to the kinds of contents suitable for report in our public language. It is this last virtue of calibrational content which comports most elegantly with the idea of a changing role of representation in cognitive science. To illustrate this, consider some recent work in the evolution of learning. In a fascinating series of simulations S.Nolfi and D.Parisi have investigated what happens when evolution is allowed to work on complex systems in which one component acts as a kind of trainer/ teacher for another. Nolfi and Parisi modelled 'organisms' consisting of two linked connectionist sub-networks. The two sub-nets both begin with random weights and share an input signal indicating the angle and distance of the nearest food source in a small grid-world. One sub-net (the 'teacher') would then feed target output patterns to the other sub-net (the 'student') whose task was to learn (using the familiar-though admittedly biologically unrealistic-backpropagation learning algorithm) the mapping specified by the combination of the input (common to both student and teacher) and the internally (teach-net) generated target output. These target outputs were to take the form of motor commands i.e. the task was to learn to respond to the sensory input data by issuing a command whose effect is (ideally) to move the simulated organism into the food-containing square on the grid, at which point the food is counted as consumed. Food items were placed randomly and re-distributed at the start of each simulation. As it stands, of course, the set-up is hopeless. It is a genuine case of the blind leading the blind. For since the weights in the teacher sub-net (which do not change during the organism's lifetime) are random, the target outputs it sends to the student will be chaotic relative to the task. The student net will learn this mapping, whatever it is, but it is unlikely to constitute any advance in the organisms overall eating skills which will remain at the level of chance. But what will happen if we allow an analogue of genetic evolution to operate on a population of such organisms? To find out, Nolfi and Parisi created an initial population of 100 organisms (teach-net/student-net pairings) each with (different) random weights in the two subnets. Each organism was allowed to 'live' for a fixed time (5,000 movements in the grid-world). At the end of that time, the number of successful ingestions was counted and the twenty organisms who had done best were each used to create five 'offspring' viz. five close copies of the initial (note, initial - this is not a Lamarkian simulation-) weights of the organism, with minor mutations (four weights changed at random). A new generation of 100 organisms was thus produced and the procedure repeated. After some 200 generations very successful eating was being achieved. This success, however, depended heavily on individual learning. The student nets of the successful organisms, at birth, were hopeless. More interestingly, the teacher nets themselves, if allowed to DIRECTLY control motor output at birth, did significantly worse than did their associated student-nets after lifetime learning. So evolution had not simply hard- wired the optimal solution into the teacher sub-net leaving it merely to transfer it to the student via back-propagation training. And most interestingly of all, if the weights in the student sub-net associated with a given teacher sub-net were randomized, the whole system failed to learn useful behaviours. So the weights in the student net, although of no direct use in solving the problem at birth, are still crucial in some way to the overall organism's ability to LEARN TO solve the problem. The way to understand what has happened here is to see that the weights in the student net 'are not selected for directly incorporating good eating behaviours ... but they are accurately selected for their ability to let such a behaviour emerge by life-learning' (Nolfi and Parisi (1991) p. 10). Unpacking this, we may note that first, the weights in the student net will have been selected so as to afford a good (local minima avoiding) starting place for learning. And second, that since the teach-net and the student-net have co- evolved, the teach-net may be in a somewhat extended sense calibrated to the particular initial location of the student in weight space i.e. it can generate training signals which reflect the specific needs of that student-net - the one whose initial weights are thus and so. In a final twist, Nolfi and Parisi repeated the experiment but this time allowed each sub-net to pass a training signal to the other (thus, in this simulation (unlike the previous one) the weights in the teach-net were THEMSELVES changeable by back propagation learning during the organism's individual lifetime). Once again good eating behaviour evolved. But this time neither sub-net was clearly acting as teacher and both were completely hopeless at birth. What we see in this final simulation is thus the power of evolution to select cleverly co-evolved complexes of networks whose joint activity, GIVEN processes of individual learning dependent upon realistic environmental inputs, is able to determine successful mature behaviour. Suppose we then sought to understand the computational roots of such success by seeking a mapping, in terms of translational content, between the inner representational resources of some highly complex version of such a system and the daily fodder of human language. In all likelihood, we would fail dismally. We would fail to illuminate the process of learning as the systems did not proceed by performing quasi-linguistic operations on a rich initial representational base. Instead, if we ask what exactly is the content of the innate knowledge embodied in the initial weights of the evolved nets, we find ourselves talking rather of one net's knowledge of the other's location in weight space, of initial positions which avoid local minima, etc. etc.: i.e. we are confronting a form of innate knowledge which is fully independent of the usual resources of daily language. This is not even calibrational content in any simple sense, as the states do not directly track features of the external environment. Instead, it is a kind of inward-looking calibrational content; it concerns the delicate evolved harmonisation of the resources of a complex of sub-networks evolved so as to yield successful learning given realistic input data. The force of this observation is not, I believe, restricted to the understanding of learning. For once we see how alien and inward-looking the initial representational base may be, we must begin to wonder whether the mature system likewise depends for its success on the exploitation of multiple delicately and probably non-intuitively harmonised (see note 2) inner resources. That is, we must begin to wonder whether the attempt to understand even the mature information-processing of the human brain will finally revolve around a body or bodies of representations whose contents are in any way well depicted by sole and direct reference to the kinds of external states of affairs by means of which we fix the meanings of public language terms. Perhaps it will not. Perhaps the cognitive scientific wisdom we seek will require us to treat the inner processing economy as itself a kind of ecology WITHIN which we must attempt (at times) to determine crucial calibrational contents. The useful notion of calibrational content may thus serve us best once it is expanded to refer not just to states of the gross external environment, but to states of the internal one as well. This notion of important but inward-looking cognitive resources can also be found in some recent work in Cognitive Neuroscience. Thus Van Essen et al (forthcoming) argue that the brain is best seen as a system evolved to treat information as "an essential commodity" and that this in turn leads to the development of very specialized systems whose task is not to represent the external world but rather to control the nature and flow of information within the system. Van Essen at al go so far as to postulate the existence of 'control neurons' whose task is to modulate the connectivity between other groups of neurons and hence to ensure that the right information arrives in the right place at the right time, thus gleaning maximal benefit from information stored anywhere in the system. These inward-looking control resources might enable the flexible linkage of a multiplicity of distinct modules and hence result in a more plastic and efficient system. But they would clearly NOT be well understood in terms of any external representational role. I conclude that the present role of representation in Cognitive Science is indeed due for a full re-evaluation. Considered as the content-bearing states of cognitive subsystems, such internal representations as we are eventually driven to posit may turn out to bear rather alien (non-folksy) and at times purely inward-looking contents. In addition there need be no single central code and content-fixation may thus depend on calibrational rather than translational techniques. Finally, representations thus reconceived will not be taken for granted as a necessary innate bedrock upon which Cognitive Scientific explanation has only to build. Representational states will be seen as much the object of Cognitive Scientific explanations as their foundation. P.M. Churchland's radical belief that the explanatory apparatus of future Cognitive Science will owe little or nothing to the "sentential categories of current common sense" (P.M. Churchland (1989) p.177) is thus one I fully and unambiguously endorse. It is a belief whose full implications we are only beginning to glimpse. 3. Blunt Instruments and Augmentative Materialism. Cognitive Scientific theorizing about mind, I have conceded, will indeed very probably depart radically from the vocabulary and assumptions of daily mentalistic talk. To the extent that it does so, must we view such talk as at best amounting to the opportunistic use of a rather blunt instrument? Must the folk vision be, if not exactly eliminated, at least relegated to the status of a one-dimensional projection of much higher-dimensional internal cognitive states, or a 'partial and unpenetrating gloss on a deeper and more complex reality' (P.M.Churchland (1989) p.7)? The view I have been developing (section 1 above, Clark (1989)) depicts it instead (and here I obviously follow e.g. Dennett (1978) (1987)) as offering a different kind of information. Thus suppose someone were to complain of the simple thermometer that it is flawed because it is just too blunt an instrument to reveal the precise nature of an illness. It would be fair to reply that that is not its job. Its job is more global, and more mundane. It is supposed, amongst other things, to alert the non-medically knowledgeable user to a general condition. Likewise, as I suggested in section 1, we may see folk psychology as a device whose purpose is to inform us of the general skills,states of knowledge and motivations of other agents. For most social and daily purposes, we care only e.g. whether so and so believes the film will begin at 10 p.m. Should they instead believe it begins at 11p.m., that fact (whose internal roots may be distributed and fragmentary, involving several possibly competing or inconsistent kinds of information stored in disparate inner resources) will explain and predict their lateness. It is this global level of detail which folk psychology is adapted to provide. What P.M.Churchland sees as an unpenetrating gloss is thus fruitfully reconceived as an effective social adhesive. One common response to such claims is to castigate them as placing too few constraints on the acceptability of the folk 'theory'. Thus in both his (1981) paper and in subsequent pieces (e.g. the commentary on Dennett reproduced in his (1989) p.125-127) P.M.Churchland accuses this type of defence of folk psychology as being too powerful. If we restrict the commitments of some discourse to the correct plotting of a few global properties which are revealed in gross behaviour, then we could defend e.g alchemy or the astronomical theory of nested crystal spheres, as acceptable. This worry deserves more attention than I have space to give it here. (For an extended treatment see Clark (1993)) But I would note that depicting the role of the folk discourse in the way I have does not commit us to the claim that the discourse makes NO assumptions which concern inner facts. My claim is just that one kind of inner fact which it does NOT concern is the question of whether we do or do not rely on some kind of sentential inner code. To give a single example, it is commonly assumed that some of my actions and verbal outputs draw on stored information concerning my previous experiences. But if it were discovered that I was a pre-programmed Giant Look-up Table with a distinct stored output for every possible input (in every possible sequence of inputs!) then that assumption would be undermined. It would then be unintelligible to suppose that my present action or utterance was caused by the stored trace of a previous experience, and it might reasonably be judged that I was not, after all, a proper object of folk psychological descriptions. The particular example is unimportant. The point is just that the stress on global behaviour does not make the folk discourse immune to inner evidence: it just makes it immune to the wrong kinds of inner evidence. The more positive vision of the nature of folk psychology rehearsed above accords surprisingly well with P.M.Churchland's own favoured model of explanatory understanding. The model in question is one in which explanatory understanding is depicted as involving prototype activation. Thus the basic form of representation in mature connectionist systems, and probably in brains, is, P.M Churchland argues, prototype-involving. What this means is that a neural network learns to use the resources of its high dimensional weight space so as to generate partitions in activation space corresponding to the features it needs to distinguish to perform a set task. Each partition separates off a sub-volume of activation space such that activity defining the centre point of the space indicates the presence of a prototypical instance of the feature (see P.M.Churchland (1989) ch.6,9,10). These prototype- style representations are best seen, I suggest, as bearing contents which are (see section 2) calibrationally determined. One appealing fact about them is that the objects of such calibration can be arbitrarily complex. To repeat an example already used, activity in the centre of one such space might be assigned the calibrational content that the mouse is eating sesame seeds. Explanatory understanding thus consists , P.M. Churchland suggests, in the activation of correct, perhaps highly abstract and complex, prototypes (see P.M.Churchland (1989) ch.10). Endorsement of such a model led P.M.Churchland to amend one of his own earlier views. Whereas in his (1981) paper he depicted our individual folk psychological understanding as 'consisting of an internally stored set of general sentences' (P.M.Churchland (1989) p.112) he now sees it as consisting in a body of stored prototype-style representations. That is, he no longer views our own internal encoding of the folk wisdom as itself stored in sentential form. Instead he expects complex partionings reflecting 'typical configurations of desires, beliefs, preferences and so forth' ( P.M. Churchland (1989) p.124). The folk framework could still mislead, he goes on to insist, as such configurations may offer only a shallow and inadequate vision of the inner roots of action. But the internal representation of the (putatively flawed) folk theory need no longer itself be conceived as sentential. The worries about shallowness we have already rebuffed; folk psychology is not shallow so much as global. Remove the taint of shallowness and we are left with a nice model of why the folk discourse, sententially couched, actually works. It works because the belief/desire (etc.) sentences successfully evoke complex non-sentential inner representations of (on our account) global behaviour patterns and trends. What P.M.Churchland says in partial defence of deductive-nonological explanation now applies wholesale to folk psychological explanation. I quote: "What a well-turned deductive-nonological argument certainly can do is successfully evoke explanatory understanding in the hearer by provoking activation of the relevant prototype.... D-N arguments are therefore entirely appropriate things to exchange in a great many explanatory contexts....... " P.M.Churchland (1989) p.224. Ditto, I suggest, for the sentential repertoire of folk psychology. Propositional attitude talk,exchanged between competent speakers of a natural language, will reliably evoke a host of inner (prototype encoding) resources whose cumulative effect is to appraise the hearer of useful knowledge and expectations concerning the likely overall behaviour patterns and responses of some other individual. In addition, and to close on a truly positive note, this overall conception lays our folk psychological understanding nicely open to the advances of what I shall dub an augmentative materialism. Augmentative materialism allows that scientific advance may indeed contribute to the enrichment of the knowledge which the folk psychological sentences reliably evoke. It may do this by e.g. adding some broad conception of the varied panoply of types of internal event which may be causing a behaviour or by expanding the repertoire of types of behaviour pattern which can be reliably identified. The process I have in mind is thus one in which the likely impact of scientific advance is to endorse and enrich the explanatory understandings reliably evoked by the folk discourse. By contrast, the kinds of scientific discovery which would genuinely undermine such discourse are, once we have disposed of the red herring of inner sentences, just the unworkable and bizarre imaginings of philosophers: pure look-up table organizations, radio controlled puppets, chance behaviours of chaotic matter etc. etc. But these are marginal phantasms whose shadowy presence is unsubstantiated by the actual conjectures of working connectionists and neuroscientists. 4. Conclusions: Provoking Prophets. The exposure of folk psychology to scientific refutation has, I have argued, been somewhat exaggerated. The single most potent source of such exaggeration is the image of the folk discourse as targeted, for some peculiar reason, on SUB-STATES of organisms. If it is instead seen as commenting on overall skills and bodies of knowledge, the argumentative dialectic is significantly altered. But whatever the fate of the speculations regarding folk psychology, the profound conceptual impact of the scientific advances upon which the Churchlands draw is real enough. We have indeed reached some kind of crossroads in our conception of the nature and form of explanation in Cognitive Science. Where once the starting point of such explanation was a conception of representation quite closely modelled on our experiences with language,proof theory and artificial grammars, we now reach a point where the nature and role of representation in the brain is almost entirely up for grabs. Several factors contribute to this healthy but challenging state of affairs. We have existence proofs of the possibility of powerful yet non-sentential modes of storing and exploiting information. We are glimpsing the possibility of models of learning and representational change which are neither constituted by nor bounded by the expressive power of any pre-existing symbol system. And in exploring the relations between evolution and individual learning we are beginning to see in concrete detail the role of unexpectedly inward-looking kinds of knowledge in promoting successful learning and the more efficient exploitation of information. Both the explanatory targets and the explanatory tools of Cognitive Science are in a state of flux. As both participants in this new ongoing work, and philosophers pursuing a unified vision of mind and its place in nature, the Churchlands' critical contribution to this emerging reevaluation is unique and crucial. Long may they provoke. NOTES 1. But see Ramsey,Stich and Garon (1991) for a valiant attempt to generate a non-obvious contradiction. I reply to their argument in Clark (1990-b) 2. A nice example of this was discovered by D.Ackley and M.Littman during some experiments aimed at determining the nature of the interactions between individual learning and evolution in a group of simulated organisms. The simulated organisms involved two neural nets, one of which produced actions and one of which evaluated resulting situations as good or bad. These evaluations were then one source of information used by the organism to modify its behaviour. Amongst several interesting results, the authors found that at a late stage in evolution some of the evaluations made by the 'judgement' net became increasingly inappropriate. Whereas in the earlier stages of evolution the judgement net had evolved so as to (correctly) judge situations involving the proximity of predators as a bad thing, it now seemed to judge that the proximity of predators was a good thing! Yet the organisms embedding these 'crazy' evaluations remained quite fit and successful (at least for a time). Why? The reason was that the action net had grown so well adapted over evolutionary time that it did not let them stay near predators for long, and so no overall damage resulted from the misjudgment. They named this effect 'Shielding'. See Ackley and Littman (1992). BIBLIOGRAPHY. Ackley, D. and Littman, M. (1992). Interactions between learning and evolution. In C.Langton, C.Taylor, O.Farmer and S.Pasmussen (eds.) ARTIFICIAL LIFE II SANTA FE INSTITUTE STUDIES IN THE SCIENCES OF COMPLEXITY vol.10. (Reading, Mass: Addison-Wesley). Churchland, P.M. (1979). SCIENTIFIC REALISM AND THE PLASTICITY OF MIND. (Cambridge University Press: Cambridge.) Churchland, P.M. (1981 ). Eliminative materialism and the propositional attitudes. JOURNAL OF PHILOSOPHY 78 no.2, p.67-90. Reprinted in P.M.Churchland (1989). Churchland P.M. (1989). THE NEUROCOMPUTATIONAL PERSPECTIVE. (Cambridge, Ma.: MIT/Bradford Books.) (1981 ). Churchland, P.M. and Churchland, P.S./ Functionalism, qualia and intentionality. PHILOSOPHICAL TOPICS 12, no.1, 121-145. Reprinted in P.M.Churchland (1989). Churchland, P.M. and Churchland, P.S. (1990). Stalking the wild epistemic engine. In W.Lycan (Ed.) MIND AND COGNITION: A READER. (Blackwell; Oxford) p.300-311 . Churchland, P.S. (1978). Fodor on language learning. SYNTHESE 38 no.1, 149-159. Churchland, P.S. (1980). A perspective on mind-brain research. JOURNAL OF PHILOSOPHY 77, no.4, 18-207. Churchland, P.S. (1986). NEUROPHILOSOPHY: TOWARD A UNIFIED UNDERSTANDING OF THE MIND-BRAIN. (Cambridge MA: MIT Press) Churchland, P.S. and Sejnowski, T.J. (1990). Neural representation and neural computation. In W.Lycan (Ed.) MIND AND COGNITION: A READER. (Oxford: Blackwell) p.224-251. Churchland, P.S. and Sejnowski, T. (1992). THE COMPUTATIONAL BRAIN (Cambridge, MA:MIT Press) Clark, A. (1989). MICROCOGNITION: PHILOSOPHY, COGNITIVE SCIENCE AND PARALLEL DISTRIBUTED PROCESSING. (Cambridge MA: MIT Press.) Clark, A. (1 991-a) Radical ascent. PROCEEDINGS OF THE ARISTOTELIAN SOCIETY, Supp.Vol.65 p.211-227. Clark, A. (1991-b). In defence of explicit rules. In W.Ramsey, S.Stich and D.Rumelhart (Eds.) PHILOSOPHY AND CONNECTIONIST THEORY.(NJ: Erlbaum) Clark, A. (1990-a) Connectionism, Competence and Explanation, BRITISH JOURNAL FOR THE PHILOSOPHY OF SCIENCE 41 pp.195-222 Clark, A. (1990-b) Connectionist Minds PROCEEDING OF THE ARISTOTELIAN SOCIETY vol XC 83-102 Clark, A. (forthcoming-a). ASSOCIATIVE ENGINES: CONNECTIONISM, CONCEPTS AND REPRESENTATIONAL CHANGE. (Cambridge, MA: MIT Press) Clark, A. (1993). Is Mind a Scientific Kind?. Paper presented to MIND AND COGNITION: AN INTERNATIONAL SYMPOSIUM, Academia Sinica, Taipei, Taiwan, May 1993. Dennett, D. (1978). BRAINSTORMS: PHILOSOPHICAL ESSAYS ON MIND AND PSYCHOLOGY. (Cambridge,MA.:MIT Press) Dennett, D. (1987). THE INTENTIONAL STANCE. (Cambridge,MA.:MIT Press) Ellis, A. and Young, A. (1988). HUMAN COGNITIVE NEUROPSYCHOLOGY. (London; Erlbaum.) Fodor, J. (1975). THE LANGUAGE OF THOUGHT. (Crowell; New York.) Fodor, J. (1987). PSYCHOSEMANTICS: THE PROBLEM OF MEANING IN THE PHILOSOPHY OF MIND. (Cambridge,MA: MIT Press) Horgan, T. and Woodward, J. (1990). Folk Psychology is here to stay. In W.Lycan (Ed.) MIND AND COGNITION: A READER. (Oxford: Blackwell p.399-420.) Humphreys, G. and Riddoch, M. (1987). To SEE BUT NOT TO SEE: A CASE STUDY OF VISUAL AGNOSIA. (London: Erlbaum) Nolfi, S. and Parisi, D. Auto-teaching: networks that develop their own teaching input. Technical Report PC1A9103, Institute of Psychology, CNR,Rome. Oaksford, M. and Chater, N. (1991) Against Logicist Cognitive Science MIND AND LANGUAGE 6, no 1, 1-38 Ramsey,W., Stich, S. and Garon, J. (1991) Connectionism, Eliminativism and the Future of Folk Psychology in W. Ramsey, S. Stich and D. Rumelhart (eds) PHILOSOPHY AND CONNECTIONIST THEORY (Erlbaum) Ryle, G. (1949) THE CONCEPT OF MIND (Hutchinson) Shallice, T. (1988). FROM NEUROPSYCHOLOGY TO MENTAL STRUCTURE. (Cambridge: Cambridge University Press.) Van Essen, D., Anderson, C. and Olshausen, B. (in press) Dynamic Routing Strategies in Sensory, Motor and Cognitive Processing, in C. Koch and J.Davis (eds) LARGE SCALE NEURONAL THEORIES OF THE BRAIN (MIT Press) Van Gelder, T. (1993) The Distinction between Mind and Cognition paper presented to MIND AND COGNITION: AN INTERNATIONAL SYMPOSIUM at the Institute of European and American Studies Academia Sinica, Taipei, Taiwan, May 1993 Warrington, C. and McCarthy, R. (1987). Categories of knowledge: further fractionations and an attempted integration. BRAIN, No.110, pp.1273 - 1296.