Rogers & McClelland's (R&M's) précis target article represents an important effort to bring explicit computational models to bear on questions of concepts and categorization; the sheer breadth of their demonstrations cannot fail to impress.
What, however, do R&M's demonstrations, presented here and in the book (Semantic Cognition, Rogers & McClelland Reference Rogers and McClelland2004), really show? R&M's ambition is threefold: to reinterpret theory-based approaches to concepts, to suggest that domain-specific knowledge emerges solely from general learning principles, and (implicitly) to undermine symbol-manipulation in favor of parallel distributed processing (PDP)-style connectionism.
The success of the enterprise lies in the simulations, and here we have serious reservations. Although the models seem to address some core phenomenon in concepts or conceptual development, few fully deliver what is promised.
Take, for example, “category coherence” (sect. 2.2. of the target article) R&M's ostensible goal is to explain why “some groupings seem more natural, intuitive, and useful for the purposes of inference than others” (sect. 2.2, para. 1), but no matter how well the models track such correlations, they never really get to the crux of the matter: causation. Some properties are correlated in the world by accident, and others because of causal relations, a fact that is apparent even to young children (e.g., Greif et al. Reference Greif, Kemler Nelson, Keil and Gutierrez2006); and people's generalization over different sets of “coherent” properties is powerfully mediated by their understanding of causal relations (Rehder Reference Rehder2003; Rehder & Kim Reference Rehder and Kim2006); R&M's model, in contrast, literally cannot represent the difference between a correlation of particular strength that is causal and one that is not.
More broadly, the demonstrations they present are (1) seriously underconstrained, (2) importantly incomplete, and (3) psychologically implausible.
1. Constraints: Each time R&M introduce a new phenomenon, they also introduce a new model: In the five core models of the target article, there are five different training regimes, five different architectures, five different dependent measures – and no effort at reconciling the differences.
2. Completeness: Although the current models excel at learning complex correlations between features, they fail to represent abstract operations over variables, structured representations, and contrasts between individuals and kinds; and it is not clear how well they can do any of these things in principle (Marcus Reference Marcus1998a; Reference Marcus1998b; Reference Marcus2001). In consequence, common two-place predicates such as X is a sister of Y, or P is a parent of Q cannot be properly represented in a fully generalizable way (Marcus Reference Marcus2001); complex notions such as lectures about movies versus movies about lectures cannot be represented without a combinatorial explosion of input nodes; and the models cannot represent basic distinctions such as the difference between the fact that dogs in general have four legs and that some dog in particular has three legs (Marcus Reference Marcus2001).
3. Psychological plausibility: Virtually all PDP models face certain fundamental problems, such as the slowness of their learning (here, acquiring a single fact can take thousands of trials), but the present models face a special set of problems, in terms of the way the learning task itself is defined. The experience of a real child can be thought of as a series of learning episodes that pertain to particular entities and whatever happens to be observable in a given moment: A child might see a particular dog, note some properties of that dog, and then update his or her internal representations. In R&M's framework, the learning experience is entirely different: In part because the model lacks a type-token distinction, the learning regime is designed such that the model never experiences any particular dog; all its experiences consist, instead, of pre-digested lessons about the properties of all dogs, whether or not those properties might be plausibly observed in any given moment. The model is not told “on a particular occasion, I see a dog, and on that occasion the dog is barking,” it is told dogs can bark; moreover, it is given all such facts simultaneously. In Figure 1, the model is told – in one go – that all canaries can grow, move, fly, and sing. Real children rarely have it so easy; at any given moment a child might see a canary that is in the act of flying, without seeing it sing, or see it sing without seeing it fly; the child has to infer that what he or she sees at one moment doesn't preclude other possibilities later. The challenge of putting together observed and unobserved properties – which some might call the heart of categorization – is not adequately addressed. (Likewise, since the model trades entirely in abstractions rather than specific experiences, there is no way to capture, e.g., the human intuition that natural kinds have essences that can persist through drastic transformations; see Keil Reference Keil1989.)
Still more broadly, perhaps the greatest problem with the current work is not so much the details as the enthusiastic reports of apparent success without comparable concern for limitations. Problems raised before (e.g., Gentner & Markman Reference Gentner, Markman, Hanson, Cowan and Giles1993; Marcus Reference Marcus1998a; Reference Marcus1998b; Reference Marcus2001; Pinker & Prince Reference Pinker and Prince1988), receive scant mention, and competing models such as Anderson and Betz (Reference Anderson and Betz2001), Rehder (Reference Rehder2003), Love et al. (Reference Love, Medin and Gureckis2004), Nosofsky (Reference Nosofsky1986), and Kruschke (Reference Kruschke1992) are never seriously discussed. There is no test, for example, of whether exemplar or ACT-R models would exhibit the same patterns, and hence no way of parceling out what predictions are unique to the present architecture and which would follow from any system that was sensitive to intercorrelations.
What we are left with? The simulations do acquire complex, learnable correlations between properties, but do not show that such intercorrelations alone suffice for human reasoning. In a true account of human reasoning, correlations may well serve as the input to cognition, but the end product is often far richer; we humans don't just notice contingencies, we seek to understand them. When a child sees that animals in cold climates tend to have heavy fur, she doesn't just note the data; she asks why, and that is one thing PDP models just cannot do.
Finely tuned statistical engines of the sort discussed here may well play some role in our conceptual understanding of the world (Keil Reference Keil, Byrnes and Gelman1991b; Marcus Reference Marcus, Broeder and Murre2000), but if the current work serves as a guide, such machinery seems unlikely to suffice on its own.
