Classicalism and Cognitive Architecture

This paper challenges the widely accepted claim that "classical" cognitive architectures can explain the systematicity of cognition (Fodor & Pylyshyn, 1988). There are plausible ways of rendering more precise the systematicity hypothesis (as standardly formulated) in which it is entailed by classical architectures, and other plausible ways in which it is not. Therefore, it is not a determinate issue whether systematicity is entailed, and hence explained, by classical architectures. The general argument is illustrated in a particular domain, the systematicity of deductive inference. In the case of the capacity to carry out the inference modus tollens, the systematicity hypothesis can be made precise in two ways, one entailed by classical architectures, another which is not. Further, the latter, but not the former, accurately describes the actual empirical phenomenon. Put another way, the clumps that these deductive inference capacities come in are not the clumps that are entailed by classical architectures. Therefore, in this area at least, systematicity considerations count against the classical conception of cognitive architecture. In their well-known paper Connectionism and Cognitive Architecture, Fodor & Pylyshyn (1988) argued that connectionism cannot constitute a viable alternative to the so-called "classical" conception of cognitive architecture, on the grounds that, unlike classical architectures, it cannot provide any explanation of the pervasive empirical phenomenon of systematicity. Therefore, the only proper role for connectionism is to investigate how classical architectures might be implemented. This argument—described by Pylyshyn as one of only two major arguments for the classical approach (Pylyshyn, 1989)—has prompted much connectionist modeling and philosophical debate. The main point of contention has been whether connectionism can deliver an adequate non-classical explanation of systematicity. Most contributors to these debates have paid little or no attention to the alleged empirical phenomenon of systematicity itself, and to whether classical architectures can in fact explain systematicity. Here we take a very different approach. We will argue that, according to Fodor and Pylyshyn's own standards of explanation, classical architectures presently cannot be claimed to explain systematicity, and further, that there is reason to believe that classical architectures cannot explain systematicity. Consequently, there can be no argument from any alleged failure of connectionism to explain systematicity to the superiority of the classical approach. Indeed, if anything, systematicity considerations currently count as an argument against the classical conception. What Is It To Explain Systematicity? Fodor & Pylyshyn place a strict constraint on what counts as genuine explanation. For an hypothesis H to explain some phenomenon S, H alone must entail S. This constraint is essential to their dismissal of connectionism. They imagine a defender of connectionism building systematicity into a particular connectionist model and claiming, on this basis, that connectionism can indeed explain systematicity. Fodor & Pylyshyn reply that this would be insufficient. Since all natural cognitive systems are systematic, systematicity must follow from the basic nature of architecture alone, and not merely be compatible with the architecture (p.50). Consequently, if the classical conception of cognitive architecture explains systematicity, it likewise must entail systematicity from the basic nature of the architecture alone. The classical conception, expressed as an empirical hypothesis, is CA: All natural cognitive systems contain (a) mental representations with combinatorial constituent structure and compositional semantics, and (b) mental processes that are sensitive to the combinatorial structure of the representations. This hypothesis must entail the alleged empirical phenomenon, which is that S: All natural cognitive systems are systematic. This entailment is neither immediate nor obvious. Whether it goes through at all depends, among other things, on what the empirical phenomenon S actually is, and this in turn depends on what the concept of systematicity is. Until systematicity is adequately clarified, we cannot know whether classical architectures explain it. What Is Systematicity? Systematicity appears nowhere in the cognitive science literature before Connectionism and Cognitive Architecture: not as an empirical hypothesis, nor as a concept; not even the term.1 It is therefore surprising 1 Productivity, of course, occurs frequently in the literature, and productivity is one component of systematicity. However, a concept does not exist merely because one component of it exists. Note, moreover, that Fodor & Pylyshyn explicitly decline to rely on productivity for the force of their argument; it is only the other components of systematicity that matter. that Fodor & Pylyshyn give no clear, succinct and precise definition of the concept, or description of the empirical phenomenon. Careful extrapolation from the scattered hints and definitional tidbits they do provide leads to the following empirical hypothesis: Systematicity according to Fodor & Pylyshyn (SFP): All cognitive systems (humans and other animals) are systematic, i.e., are such that their ability to do some things of a given cognitive type (including at least "thinking a thought" and making an inference) is intrinsically connected with their ability to do other, structurally related things of that type. This statement is as precise as Fodor & Pylyshyn get. However, it is not sufficiently precise for determining whether classical architectures explain systematicity. There are ways of sharpening SFP such that it is entailed by CA, and ways of sharpening it such that it is not entailed by CA. Therefore, it is not a determinate issue whether CA as it stands entails SFP, or systematicity in general. The remainder of this section states this argument in a little more detail. How might SFP be rendered precise? As noted, we cannot turn to the existing literature for help, for there is none bearing directly on the issue. (This may explain the otherwise curious fact that Fodor & Pylyshyn, in the body of Connectionism and Cognitive Architecture, cite no empirical literature whatsoever in support of their claim that cognition is systematic.) Consequently, we must begin from scratch. The basic idea behind systematicity is that any organism able to do one thing of a given type is able to do other, structurally related, things of that type. To render systematicity precise is to get clear on what cognitive performance types there are, and on what other things of a given type an organism would have to be able to do, if it can do some particular thing of that type. This suggests the following schema: Systematicity Schema: For every organism O, and any given cognitive performance t of type T, there is some set MO,t of "structurally related" performances such that O is capable of all and only the performances in MO,t. This schema would then be fleshed out for particular organisms and performance types in order to provide empirically applicable tests. There are many ways the Systematicity Schema might be filled out in detail for particular performance types. Some of these ways are entailed by the hypothesis that cognitive architectures are classical (CA). Some are not. Suppose, for example, that the sets MO,t were specified to be some proper, nonarbitrary subsets of the sets that would be entailed by classical architectures. (An example is given in the next section.) Then those cognitive capacities would clearly be systematic, but systematicity of this kind would not be entailed by classical architectures. for classical architectures entail that the organism is capable of performances that are not in those sets. Consequently, in the absence of any particular specification of how the Systematicity Schema is to be filled out, there is no determinate answer to whether classical architectures entail and hence explain systematicity. Further, since no-one has in fact provided such specification in any reasonable detail, no-one can presently justifiably assert that classical architectures do entail and hence explain systematicity. Think of it this way. CA entails that cognitive capacities come in clumps, i.e., are systematic. SFP asserts that cognitive capacities come in clumps. Does CA therefore entail SFP? Not necessarily. It only does so if the kind of clumps entailed by CA are the same kind of clumps picked out by SFP. So, what kind of clumps does SFP pick out? Well, that's not clear at all. On one way of reading SFP, the clumps are the same. On another way, the clumps are not. Until we've settled on a specific way of reading SFP, we just can't say whether the entailment is there. Unfortunately, Fodor & Pylyshyn didn't provide any specific reading, and nobody else has either. So, we can't now say that classical architectures do entail, and hence explain, systematicity. How Classical Architectures Do Not Explain Systematicity The previous section argued that it cannot now be said that classical architectures entail systematicity. This section argues for a different point: that in one domain at least, classical architectures do not entail the actual empirical facts of systematicity. In making this second argument, we provide a concrete illustration of the key premise of the first argument, which is that there are various ways of making systematicity precise, some which are, and some which are not, entailed by classical architectures. Fodor & Pylyshyn single out the systematicity of inference as a key component of the wider phenomenon of systematicity. It is, roughly, the idea that the ability to make some inferences is intrinsically connected to the ability to make other, logically related inferences. They offer no precise definition of the phenomenon and cite no literature in its support, but do anecdotally illustrate what they have in mind the following way: You don't, for example, find minds that are prepared to infer John went to the store from John and Mary and Susan and Sally went to the store and John and Mary went to the store but not from John and Mary and Susan went to the store. (p.48) Perhaps; perhaps not. In any case, the import of such an casual observation for the general phenomenon of systematicity of inference is entirely unclear. A more appropriate procedure is to fill out the Systematicity Schema for particular inference types. Here we discuss only one, modus tollens (A⊃B, ~B => ~A). Note that modus tollens is one of the simplest and most common of all inference types. If classical architectures fail to explain the actual systematicity of modus tollens, this will significantly undermine the claim that classical architectures can in fact explain the systematicity of inference and indeed systematicity in general. There are at least two ways to flesh out the Systematicity Schema for this form of conditional inference. One is such that the systematicity of this capacity is entailed by classical architectures. Such architectures postulate mental processes that operate on mental representations in a way that respects their combinatorial structure. Classical architectures therefore predict that (subject to resource constraints) any cognitive system that can perform any instance of modus tollens (i.e., can construct conditionals and negations, and is able to draw the appropriate conclusion) will be able to perform every instance, since all such instances have the same combinatorial structure. In particular, classical architectures predict that, since mental processes are sensitive to structure, such features of the inference instance as the content of the constituent symbols or their frequency of prior occurrence should be irrelevant, since such features make no difference to the combinatorial structure. More precisely, classical architectures entail the following systematicity sub-hypotheses. Let MO,MT be the set of inferences by substituting into the modus tollens schema (A⊃B, ~B => ~A) any symbol in the set of symbols available to O. Then we have: Systematicity of Modus tollens (SMT): Any organism O capable of performing any instance of MO,MT is capable of performing every instance of that set. (Note that this instantiation of the Systematicity Schema has been simplified by assuming that the set M does not depend on any particular performance t.) Hypothesis SMT expresses in precise terms one way in which cognitive capacities can be said to come in clumps. It is entailed by the hypothesis that cognitive architectures are classical in form. Does it accurately describe the kind of clumps that cognitive capacities actually come in? Conditional inference has been the target of much psychological investigation. Though this investigation was not specifically directed at evaluating any systematicity hypothesis, it does shed a certain amount of light on the issue. The general situation is dramatically illustrated by Table 1, from a study by Kern, Mirels & Hinshaw (1983). In this study, scientists were presented with conditional inferences of four kinds and asked whether certain conclusions followed. Some were presented inferences in abstract form (Do P⊃Q and not-Q imply not-P?) and others in concrete form (Do If Rex is a terrier, then he likes apples, and Rex does not like apples, imply Rex is not a terrier?). For current purposes, the crucial thing to notice is the disparity in performance between abstract and concrete instances. Note, for example, only 41% of scientists in one group correctly recognized the validity of modus tollens in an abstract case, whereas 69% in another recognized its validity in a concrete case. Assuming representativeness, this suggests that around 30% of people do not perform identically on structurally identical inferences, i.e., directly violate hypothesis SMT. The clumps that these people's cognitive capacities come in are not the clumps entailed by the hypothesis that they have a classical architecture. For various reasons it would be inappropriate to place too much weight on these figures alone. The moral they suggest has, however, been borne out repeatedly in numerous systematic studies. The difference between the abstract and concrete cases is that the latter have meaningful content. The critical role of content in conditional inference has been confirmed repeatedly in one of the most-studied tasks in the psychology of inference, the Wason card selection task (Wason, 1966). In the standard version of this task, four cards are laid out on a table before the Table 1. Scientist's performance on simple conditional inferences. From Kern, Mirels & Hinshaw (1983)