What's really wrong with the computational
theory of mind
Jonathan Knowles
Philosophy Department, Norwegian University of Science and Technology, 7491 Trondheim, jonathan.knowles@hf.ntnu.no
Jerry Fodor has long insisted that the so-called computational theory of mind (CTM) is 'the only game in town' when it comes to giving an account of the natural mechanisms underlying thought and cognition (at least since Fodor 1975). CTM has it that the mind, at least to some significant degree, has the structure of a digital computer: thoughts are higher level brain states with a syntactic structure mirroring their semantic articulation, and thought processes are computational operations defined over such structures, sensitive only to their formal, syntactic properties. Many arguments have been given for CTM over the years, from the need to reconcile the abstract and external nature of thought content with its causal relevance, to the need to explain natural language learning and understanding (cf. Fodor op. cit., 1980, 1987). In recent years, the main focus of the debate around CTM has been its advantages and disadvantages relative to an alternative conception of cognitive architecture broadly known as connectionism. Connectionism involves less commitment to the literally computational nature of thought, but is most directly motivated by the desire to give an architecture with greater fidelity a) to the biological structure of the brain and b) to the fluid and flexible nature of cognitively-driven behaviour. In reality, connectionist theories of cognition are computer models that are little more than inspired by the structure of the brain, but they nevertheless involve a clear rejection of the central idea of CTM: that computational operations are defined over entities with a structure isomorphic to that of the semantic interpretation of the system. Fodor and his sympathizers have retorted that connectionism thereby cannot do justice to the nature of thought in respect of its systematic and rational articulation: Only a classical (i.e. Turing/von Neumann) architecture can give an explanation as to why our thought processes are systematically interanimated with one another, and also respect the phenomenon of logical, rather than merely associative transition between such thoughts. For Fodor, connectionist models of cognition are at best implementation theories of classical architectures.
In his most recent statement on CTM, Fodor is less sanguine about the prospects of CTM as a full account of human cognition (Fodor 2000). However, this is not because he has in any way retracted from the 'only game in town' motto. He argues - building on a theme developed in earlier work (Fodor 1983) - that CTM is ideally suited to modelling thought processes that are essentially local in their operation, but that higher level human cognition, in which belief, typically, is fixed against a large array of background assumptions, is quintessentially non-local. Connectionism, however, cannot solve this problem either, since - having abjured the idea of thoughts identifiable across different cognitive contexts - it cannot even state it. Fodor is more concerned with what he calls the 'New Synthesis' in cognitive science, which incorporates Chomskyan neo-rationalism (i.e. nativism) about thought content, something like his own CTM concerning thought processes, and an adaptationist-inspired massively modular conception of the coarse-grained structure of cognition (the massive modularity thesis, or MM) - according to which all cognition is divided into circumscribed, domain specific processing units. Now Fodor is no enemy of modularity in itself; indeed he recommends concentration on modular systems in cognitive science insofar as, in being local in their operation, they are more likely to be amenable to modelling by CTM. However, the extreme view of MM is, he thinks, theoretically incoherent. This is because modules require a kind of filtering of inputs to arrive at the kind of restricted range of data upon which they act. This filtering process will ex hypothesi (i.e. by MM) also be a modular process, which in turn will require a further modular filtering system, and so on and so forth. Thus a regress is set up that can only be terminated by a non-modular system. MM is thus self-refuting - yet is the only architecture that would have held out hope of understanding cognition in its entirety. Central, domain-unspecific cognition is thus essentially a mystery rather than a scientific puzzle, to employ a distinction due to Chomsky - even though CTM, and the idea of a syntactically structured internal code that goes with it, remains untouched as far as it goes.
Fodor's arguments against MM have not gone uncontested (cf. Collins 2005, Carruthers 2003). However, the disagreement I have with Fodor does not lie here. For Fodor, MM is a forlorn thesis, but at least insofar as thought is modular, CTM stands ready to provide an account of it; and even in the absence of moduliarity, CTM is still `the only game in town', and thought quite generally has to be seen as underpinned by internal, structured vehicles. In my view, CTM can be seen to be flawed once one makes explicit the nature and extent of its applicability in this way. Given Fodor's scepticism with respect to MM and to the possibility of a fully mechanistic account of central cognition, both of which I regard as reasonable, connectionism has in fact greater promise as an architectural theory within cognitive science than Fodor gives it credit for, indeed, probably greater than CTM - both in relation to modules and central cognition.
Consider first modules in the light of what I take to be Fodor's master argument for CTM. According to this argument, what the digital computer provides is an account of thinking that ensures that semantic transitions between thoughts are respected in the causal functioning of the system in which they are realised. An important lemma here is that proof theory provides a formal analogue to the semantic theory of logical consequence; but a computer can do proof theory, hence, if our brain is a computer, we can understand how we draw logical conclusions from our thoughts. But not otherwise, thinks Fodor - in particular, a connectionist machine will at best be able regularly to associate a thought of the kind q with p & q, not to ensure that q is a consequence of p & q. Similar considerations are marshalled against the idea that connectionist machines can explain the fact that a thinker who can think Fa (viz. John is bald) and Gb (viz. Jane is beautiful) ipso facto is able to think thoughts of the form Ga and Fb (cf. Fodor 1987: Appendix, Fodor & Pylyshyn 1988, Davies 1991).
The problem with this argument is simply that these kinds thought processes are not at all clearly applicable to modules - just where Fodor thinks CTM has best prospects for unmitigated success. Modules are not ipso facto subject to the kind of systematicity requirement adumbrated, for their contents are not conceptualised by thinkers (cf. Davies 1989). Further, the transitions between representational states in modules will (typically) not be formally valid, rather they will be truth-preserving given certain specific environmental conditions obtain - for example, in the visual system, the valid operation of the algorithm for deriving information about reflectance changes in surfaces presupposes these surfaces vary smoothly (cf. Marr 1982: § 2.1). In general, it seems modular inferences will be highly context-specific in the sense of holding only given the stability of the particular environmental conditions in which the module evolved. There is thus no question of a mechanism of thought that ensures that a module's transitions are truth-preserving - that is, a mechanism implementing something like proof theory, for there is nothing like proof theory that could be implemented in these cases. This does not mean that there cannot be good empirical reasons for thinking that syntactically structured representations might be required in modules (see e.g. Margolis & Laurence 1999), but it seems clear that the very basic motivation Fodor sees for thinking mental processes must, literally, be computational cannot apply here. Moreover, once one accepts this, even arguments for syntactically structured representations will need to be answerable to considerations from lower down in the explanatory hierarchy, along the lines of what Patricia Churchland has called the co-evolutionary research strategy (Churchland 1986). Assuming this will bring in considerations concerning the structure of the brain, this will probably give some independent support to neurally-inspired, i.e. connectionist architectures.
When we turn to the central system of cognition, one might think that Fodor's master argument would have more force - for here, surely, we precisely do find necessary structure and strict logical transitions between thoughts. However, it is important to remember that, by Fodor's own lights, CTM cannot provide an account of central thought processes in their entirety. Further, though the argument for structure he and others have given is perhaps a good one in the context of the aim of giving a reductive account of thought, it is far from clear that it remains so once the context is that of having admitted that maybe no such reductive account - at least along the lines of CTM - can be given. To start with, the very idea of a mystery - of the kind we are left with in going along with CTM `half-way', so to speak - is not entirely happy: It is hard to see how we can pronounce so definitively on anything whose nature we are meant to understand so poorly. A less paradox-ridden approach would be to open up for scientific pluralism, inspired by recent developments in many other areas of science, notably biology (cf. Dupré 1993). If we open for pluralism, the idea that thoughts are to be viewed as sub-personal syntactic entities when they are, in the first instance, something quite different - presentations of the world for an individual in conscious thought (cf. Knowles 2005) - should not seem compelling. The everyday notion of representation becomes strained in converting it to talk of representations as internally interacting items, and we need good reasons for making this conversion. In the absence of the prospect of a reductive account, we would arguably do better to stick with the common sense notion, at least pro tem, and attempt to provide some kind of intra-level, non-reductive understanding of it.
At the same time, such a pluralism combines naturally with a non-reductive account of the underlying mechanisms of thought (i.e. one which does not seek an explanation in terms of internal `representations'). Moreover, it is precisely this kind of account it seems connectionism is in a good position to offer, insofar as holistic and flexible operation (inter alia) will be central desiderata. Connectionism in this role may only be an `implementation' theory, but in the dialectical context we are here considering this is not a draw-back - it will not, at least, be an implementation of a classical architecture, as Fodor has charged (cf. Knowles 2001: § 3).
In sum, given the limitations on CTM Fodor admits - of applying most naturally to modules and being woefully inadequate when it comes to central cognition - the central arguments for it lose cogency, whilst developing alternative architectural theories emerges as a more promising line of research.
References
Carruthers, P. (2003) `Massively moderate modularity' in A. O'Hear Mind and Persons, Cambridge: Cambridge University Press.
Churchland, P. (1986a) Neurophilosophy, Cambridge, Mass.: MIT Press.
Collins, J. (2005) `On the Input Problem for Massive Modularity', Minds and Machines 15: 1-22.
Davies, M. (1989) `Tacit knowledge and subdoxastic states' in A. George, ed., Reflections on Chomsky. Oxford: Blackwell.
Davies, M. (1991c) `Concepts, connectionism, and the language of thought' in D. Rumelhart, W. Ramsey, & S. Stich, eds., Philosophy and Connectionist Theory. New Jersey: Lawrence Earlbaum and Associates.
Dupré, J. (1993) The Disorder of Things. Metaphysical Foundations of the Disunity of Science, Cambridge, Mass.: Harvard University Press 1993.
Fodor, J. (1975). The Language of Thought, Cambridge, Mass.: Harvard University Press.
Fodor, J. (1980) `Methodological Solipsism Considered as a Research Strategy in Cognitive Psychology' Behaviorial and Brain Sciences 3.
Fodor, J. (1983). The Modularity of Mind, Cambridge, Mass.: MIT Fodor, J. (1987). Fodor, J. (1987) Psychosemantics: The Problem of Meaning in the Philosophy of Mind, Cambridge, Mass.: MIT Press.
Fodor, Jerry A. (2000). The Mind doesn't Work that Way, Cambridge, Mass.: MIT Press.
Fodor, J. & Pylyshyn, Z. (1988). "Connectionism and Cognitive Architecture: A Critical Analysis" in S. Pinker and J. Mehler, eds., Connections and Symbols, Cambridge, Mass.: MIT Press (A Cognition Special Issue).
Knowles, J. (2001) `Does intentional psychology need vindicating by cognitive science?' Minds & Machines 11: 347-377.
Knowles, J. (2005) Review of T. Crane The Mechanical Mind, 2nd edition. Minds & Machines 15: 259-270
Margolis, E. and S. Laurence (1999). "Where the Regress Argument Still Goes Wrong: Reply to Knowles." Analysis 59(4): 321-327.
Marr, D. (1982) Vision. San Francisco: Freeman.
Short abstract for:
What's really wrong with the computational theory of mind
Jonathan Knowles
In The Mind doesn't Work that Way, Jerry Fodor has expressed reservations about the explanatory capabilities of what he has described as `the only game in town' when it comes to an account of the natural mechanisms underlying thought and cognition: the computational theory of mind (CTM). Building on earlier themes, Fodor argues that CTM is ideally suited to modelling cognitive processing which draws on a restricted range of information - that characteristic of input systems or modules - but that it provides no or little insight into processes of a more global nature, such as the fixation of belief in central cognition. The recently popular idea that cognition might be massively modular, though in principle promising as a way forward here, is self-refuting because of the need to filter inputs to modules (the so-called input-problem). In my view, once it is conceded that it is modules CTM applies most naturally to, and that it is at best woefully incomplete as an account of central cognition, the main motivation for CTM vis à vis more recent connectionist architectures vanishes. The processing carried out by modules is typically highly context-sensitive in the sense of presupposing facts about the environment to operate reliably, i.e. modular inferences are not formally valid. Hence nothing like proof theory could be used to understand how they preserve truth - yet the need for something like proof theory is an important lemma in Fodor's argument for CTM. When it comes to central cognition, that no fully reductive account of this is in sight suggests we should instead adopt a more pluralist stance in cognitive science, seeking to understand conscious, representational thought in the first instance as such, i.e. at the personal level. Connectionist models, meanwhile, look well-suited to a providing a (non-reductive) account of the mechanistic basis of higher thought processes, but will not thereby be viewable as implementing a classical, i.e. CTM architecture, as Fodor has often charged.
What's really wrong with CTM
3 / 11-12-08
1 / 11-12-08