Please do send comments: email
Back to my home page.
The Mind Doesn’t Work That Way:
the scope and limits
of computational psychology
By Jerry Fodor (Cambridge, MA: MIT Press, 2000)
Review by John Sutton, February 2001
The Computational Theory of Mind (CTM) can’t explain "much of what’s special about our kinds of minds". Or so argues Jerry Fodor, who has defended this very theory more vigorously than most over the past 25 years. And Fodor’s not talking about consciousness, or emotion, or even about the ultimate origin and nature of meaning and intentionality, mysteries which CTM was never designed to dissolve. So what’s going on? Has Fodor finally been caught in the connectionist net? Or (God, Granny, and Turing forbid) started worrying about phenomenology or ‘embodiment’? No. In fact this short, difficult, quirky, important book is no heresy in the High Church of Classical Computationalism, but rather a series of doctrinal skirmishes with overzealous and over-popular sympathizers. It is essential reading for anyone interested in the philosophy of cognitive science.
Fodor’s target is what he calls the ‘New Synthesis’, a combination of CTM and evolutionary psychology developed in the last decade by Leda Cosmides and John Tooby, Steven Pinker, and Henry Plotkin. Fodor’s sensibilities are affronted by the relentless optimism of these writers in claiming to know "how the mind works", by the way they preach hard-headedness while bluntly neglecting complex issues in philosophy of science, and by the prospect of their politicized phylogenetic speculations coming to dominate public perceptions of cognitive science. More importantly, he thinks most of their arguments are "appalling". Since Fodor shares many of their presuppositions (being committed not just to CTM but also to some strong variety of nativism), and is as hard-headed as they come, they have found a worthy opponent.
Fodor is most effective in rejecting the pan-adaptationism of the evolutionary psychologists. He rejects three arguments for supposing that ultimately, historically, we have the cognitive capacities we do because of our genes’ selfish ‘need’ to propagate. Cosmides and Tooby claim that the methodological requirement of consistency between the sciences entails that our cognitive architecture must be an adaptation: Fodor responds, rightly, that such a priori considerations alone cannot show that psychology and evolutionary biology actively constrain each other any more than do lunar geography and the theory of cellular mitosis. Along with Pinker and Dennett, they argue that understanding the function of cognitive systems must be done by investigating their selection history: Fodor responds, persuasively, by suggesting that non-historical, synchronic accounts of function are equally illuminating, and a lot easier to come by. And along with Pinker, Plotkin, and Dawkins, they suggest that the sheer complexity of cognition means that it must have evolved under selection pressure: Fodor replies, plausibly, that the inference from complexity to adaptation isn’t inevitable, for relatively small neurological reorganizations (requiring little genotypic alteration) may have resulted in crucial and discontinuous psychological changes. This last thought is provocative, if underdeveloped: given how little we know, says Fodor, about how mind supervenes on brain, there is no reason to assume in advance that the evolution of cognition must have taken the continuous linear path which classical Darwinian selectionism requires.
Fodor’s caution about the hype around evolutionary psychology, then, is a salutory if (as he says) "jaundiced" corrective to a bewilderingly rapid consensus, and it doesn’t spring from the ignorance of biology of which he’s occasionally accused. Ontology aside, Fodor is surely right in arguing that we need much better theories of cognitive architecture, mental representation, and cognitive processes before our speculations about the evolution of cognition can be adequately constrained. It would help if Fodor offered a more explicit account of what he takes a Darwinian adaptation to be: but the possibility that much of cognition may turn out not to be an adaptation already makes some more specific debates in cognitive science about function, such as that between J.A. Hobson and Owen Flanagan on dreams, look beside the point.
Fodor does believe that, having set aside such general adaptationist claims, it’s likely that certain restricted and peripheral parts of the mind have required detailed instruction from the world, though natural selection, to acquire an innate stock of contingent true beliefs in the realms, say, of space and object perception. But he doesn’t think that the mind can be entirely composed of such distinct "modules", as the evolutionary psychologists require. The argument here is the puzzling ‘input problem’, which demands that some non-modular, general purpose mechanism classifies which inputs go to which restricted modules (pp.71-78). Since Fodor does think there are some modules, he must believe that this problem is sometimes solved. But he pessimistically notes that, even in the case of the paradigmatic module for language perception, no-one has managed to show how sensory data comes nicely carved up into the right sets of clues for it to be appropriately channelled. Another oddity here is that, in his earlier attack on Pinker’s ‘massive modularity’ thesis, in the London Review of Books, 15 Jan. 1998, Fodor’s case focussed on the other end of the process, on the ‘integration problem’: "eventually the mind has to integrate the results of all those modular computations and I don’t see how there could be a module for doing that. … If, in short, there is a community of computers living in my head, there had also better be somebody who is in charge; and, by God, it had better be me".
It’s curious, then, that this integration problem doesn’t appear in the book. The ‘post-modular’, central processes under consideration instead, those which spell trouble for CTM, are characterized as global, abductive, or context-sensitive thinking. As critics of classical cognitive science have long pointed out, humans are generally pretty good, and computers generally pretty bad, at estimating the general relevance of one thing to another. When acquiring new beliefs, we often do so not by adding them in isolation, one by one, but by harmonising them in some maximally simple way with our existing beliefs; when changing our minds, we tend to alter those beliefs which are least central to our entire web. Such judgements of relevance are part of the unspoken psychology of common sense, and, as Fodor points out, the difficulty of building in such sensitivity to the context or ‘background’ has proved the greatest embarrassment of classical cognitive science and robotics. The fact that Pinker and Plotkin simply fail to address abductive inference or the frame problem is a sign not that we know how the mind works, but that "much of the field is in deep denial".
Fodor’s diagnosis, which again will be music to the ears of long-time foes like Hubert Dreyfus (unnamed here), is that there’s an intrinsic tension between the abductive processes which characterize the most interesting features of human cognition, and CTM itself. In its classical form, on which Pinker, Cosmides, and Tooby rely even more than they explicitly avow, CTM suggests how meaningful processes can be realized in matter. Causal transactions between mental representations, according to CTM, are driven only by the local properties and arrangement of the parts of the representations. But, as this is a computational system, these physical properties are thus also syntactic properties, and the causal interactions between representations will, all else being equal, respect the semantic relations between representations in a language of thought. This, Fodor sighs, was and is "a lovely idea" for how rational processes could be mechanized: "thinking can be rational because syntactically specified operations can be truth preserving insofar as they reconstruct relations of logical form; thinking can be mechanical because Turing machines are machines".
But it’s essential to the power of this idea that only the essential, context-invariant, syntactic properties of mental representations have causal powers: and, unfortunately for CTM, many cognitive processes are causally driven by global or contextual factors which are not intrinsic or essential. So CTM can only apply, at best, within restricted modules which have a limited database of information, and which thus do not face the problem of having to search through entire belief systems. Fodor’s grim advice then, is that cognitive scientists should concentrate on informationally-encapsulated modular systems, which are minimally afflicted with globality, until someone has a good idea about how to understand abduction.
Even for readers delighted with Fodor’s attacks on evolutionary psychology, there are many puzzles and gripes in this tightly-packed and rangy book. It’s annoying, for example, that Fodor persists in refusing to understand clear connectionist proposals, by Cottrell, Churchland, O’Brien, and others, for understanding type identity of vectors by way of structural isomorphism. But let me briefly probe Fodor's consistency organ, to try to piece together the consequences of his argument for the question of the putative scope of CTM. Fodor claims here that he’s never been prone to the Pinkerish hope that CTM will prove to be a very large part of the whole story about the mind. Indeed, it is a shock to go back to the final pages of Fodor’s The Language of Thought (1975), to find him granting the importance of swathes of non-computational causal factors in psychology, including both non-cognitive physiological factors and non-computational cognitive processes like association and ‘creative’ thinking, whatever that is. There, Fodor stressed the possibility of voluntary, if sometimes indirect, control over these non-automatic processes, thinking presumably of the role of what he elsewhere calls the "central integrator guy", who does all the stuff the soul used to do. But the picture, anyway, is that there are two radically different kinds of thinking. One is ‘rational’ in the restricted sense that it preserves truth in virtue of the logical form of the representations in play: this kind of thinking takes place in modules with restricted databases, and will probably be understood fairly well with the resources of CTM. The other kind is global, abductive, analogical, and covers forms of practical reasoning which draw on much wider informational resources: this, says Fodor here, is true thinking, not just computing, is by definition not a modular process, and cannot be modelled by CTM.
But, whereas on most views a cognitive process either is or is not computational, whether it is local or global seems to be a matter of degree. Fodor doesn't develop any examples of within-module computation in sufficient detail to suggest just how much background might creep in. So the likely scope of CTM is unclear in a world in which most of our inferences are non-monotonic, defeasible by further premises. Fodor notes that CTM was never meant to apply only to valid inferences, but accepts that attempts to extend the program to cover "all sorts of inferences that are heuristically reliable though not valid (and, for that matter, all sorts of tempting but fallacious inferences)" have not been successful in practice (p.108). This at least does seem a retreat from his earlier views: in 1988, for example, Fodor and Pylyshyn described classical cognitive science (as based on CTM) as "an extended attempt to apply the methods of proof theory to the modeling of thought" and to inference in general. They were not then suggesting "that logical proofs per se are so important in human thought, but that the way of dealing with them provides a clue as to how to deal with knowledge-dependent processes in general". And certainly Fodor's notorious rationalist rhetoric used to rely on a quite general "similarity between trains of thoughts and arguments", a similarity displayed, for example, in Conan Doyle's descriptions of Sherlock Holmes' reasoning. Conan Doyle, wrote Fodor in 1985, "was a far deeper psychologist - far closer to what is essential about the mental life - than, say James Joyce (or William James, for that matter)". A mere stream of consciousness, like the associative flow of Molly Bloom's night thoughts, might be "mental causation perhaps; but it's not thinking". Well where, then, for Fodor, is thought? Down among the disciplined homunculi computing away in the language of thought, or up in the mysterious abductive central processes, with their "passion for the analogical"? Those without prior commitment to CTM may suspect that the dichotomy is false.
It was, after all, to be one of the great virtues of CTM that it showed how folk psychology could be true, how the beliefs and desires with which we predict and explain actions can be just as robustly real, just as rich in causal powers, as rivers and rocks. Since the falsity of folk psychological explanation (in, for example, an eliminativist scenario) would, for Fodor, be the greatest intellectual catastrophe in history, way overshadowing the death of God, the stakes are high. Yet, as Fodor always acknowledged, rich folk psychological theorizing is always vulnerable, always hostage to new information, so that our homespun folk-psychological generalizations must be hedged in with ceteris paribus clauses. But what the present book seems to suggest is that, at the heart of the most characteristic human cognitive processes, all else is very rarely equal. CTM may work at the periphery, in innate automatic perceptual modules, but the common humanity which this offers doesn't go far, and the desire Fodor expressed in 1988 that modularity would show how we "transcend the particularities of our training" and experience looks again to be under threat. It is not surprising, then, that folk psychology is rarely mentioned in this book. Towards the end Fodor does again endorse the bare notion of an innate theory of mind module: but this is driven more by his general scepticism about the possibility of genuine learning than by any argument that folk-psychological inference escapes the affliction of globality. If, as Fodor says, "the cognitive mind is up to its ghostly ears in abduction", and if it's for this reason, as Fodor says, that classical AI has failed "to produce successful simulations of routine commonsense cognitive competences", then the true friend of folk theory of mind might be well advised to give up classical versions of computationalism entirely.
John Sutton teaches philosophy at Macquarie University in Sydney, Australia. He is author of Philosophy and Memory Traces: Descartes to connectionism (Cambridge University Press, 1998), and co-editor of Descartes' Natural Philosophy (Routledge, 2000).
TOP OF PAGE
Updated 12 February 2001.