If Rogers & McClelland (R&M) have done what they say they have done in Semantic Cognition (Rogers & McClelland Reference Rogers and McClelland2004), then they have solved one of the most vexing problems in philosophy, cognitive science, and psychology: namely, the problem of giving a mechanistic explanation of why humans are able to have thoughts and beliefs that are about objects, states of affairs, propositions, and the like. That is, should R&M's claims hold true, so that the neural network they propose indeed acquires semantics, then they should be considered to have solved the long-standing problem of intentionality (Brentano Reference Brentano1874/1995). It is thus important to consider the question whether the neural network proposed indeed does this. That is, we need to evaluate whether the network really acquires “internal representations of objects” (Semantic Cognition, p. 69), so that it would be accurate to state, say, that the network “represents the daisy, the sunflower, the canary, and the sunfish as similar to one another” (p. 287), or that it has “Like the children, … .inferred that the object [an echidna] can move and that it has legs.” (p. 252).

Let us consider an example to see whether such claims hold up to scrutiny. We read, on page 215, that the network displayed “a tendency to extend the name ‘dog’ … to robins.” So, according to R&M, their network has learned to apply the name “dog” to some objects, namely, dogs, and now extends that name to other objects, namely, robins. Obviously, R&M cannot be taken to seriously mean this. Their network has never seen dogs or robins. Hence, it has never applied the name “dog” to any dogs whatsoever and could not possibly “extend” that name to robins, simply because there aren't any robins around to extend it to. Now, if the network cannot be said to learn to apply the concept of “dog” to dogs, or to extend that concept to robins – as we think is painfully obvious – then what can it be said to do?

Clearly, the network can be said to respond to certain input patterns with certain output patterns. What do these patterns consist of? R&M assert that the network is trained with the corpus of Collins and Quillian (Reference Collins and Quillian1969), that is, a set of concepts (“dog,” “robin”) with associated properties (“is a living thing,” “can fly,” etc.). Literally speaking, however, this is incorrect. What the network is trained with, and probed to reproduce, is a pattern of zeroes and ones that was extracted from Quillian's corpus. “Extracted by whom?,” one may now plausibly ask – after all, we are talking about semantic cognition here, and if it were the neural network that extracted these patterns, say, from its observations of dogs and robins – real or simulated – then that would at least count for something. The disappointing answer, however, is that these patterns were not extracted by the network, but were carefully put in place by R&M themselves.

A second problematic aspect of the network architecture is that the network learns to associate a label, say, “dog,” with the corresponding properties, for example, “animal,” “tail,” “bark,” and so forth. Children, when learning to make sense of the world, do exactly the opposite: They see a bunch of features, and their task is to learn that there are similarities between objects based on these features, and that, if similarity is large enough, objects should be clustered into a single concept. The network presented by R&M seems to be learning language by reading a dictionary, which is most certainly not the way children learn a language. Moreover, the network by R&M fails to capture an important aspect of categorization learning, which is the sudden of stepwise nature of that learning process (Schmittmann et al. Reference Schmittmann, Visser and Raijmakers2006).

But wouldn't some semantic cognition be necessary for the network to operate as well as it does? Certainly not. Despite the superficial resemblance of a connectionist model to the neural structures in our heads, all that the network of R&M does is to carry out a statistical categorization procedure. No semantic cognition is required for this purpose, as can be easily verified by looking at other statistical programs of this type, for instance, those incorporated in statistical packages like SPSS. It is ironic, in this respect, that R&M use clustering and multidimensional scaling programs to study the “representations” that the network “has” (e.g., see Semantic Cognition, Ch. 7), for their network is exactly such a program. Moreover, given that such wonderful statistical techniques as multidimensional scaling and hierarchical clustering are available to describe these learning processes, why not just stick to those techniques rather than use an overparameterized version of them implemented in a neural network (cf. Ripley Reference Ripley1996)? The usually quoted advantage of neural networks over statistical techniques is that they behave better in the face of noisy data; that advantage, however, is almost certainly completely lost in the R&M network, because it has localist rather than distributed representations.

We submit that the neural network of R&M does not acquire any semantics whatsoever, and as a result their theory cannot be a theory of semantic cognition. It seems to us that R&M have two options in responding to this charge, of which one is hopeless and the other absurd. The hopeless route to defuse the criticism is to attempt to show that a neural network in fact does something qualitatively different from, say, principal components analysis. This is hopeless because, despite all the fancy talk of “neurons” that “fire” to create “distributed representations” and the like, the sobering fact is that neural networks, in general, can be shown to do nothing but principal components analysis, maximum likelihood regression, and so on (cf. Hadley [Reference Hadley2000] for relationships between classical computational models and neural networks; and Sarle [Reference Sarle1994] for an overview of equivalencies between neural networks and different regression techniques and principal components analysis). Alternatively, R&M may take the other horn of the dilemma and argue that humans do nothing but carry out statistical categorization procedures in acquiring semantics. This is not logically inconsistent, to be sure, but it has some very unhappy consequences: If there is nothing to semantic cognition except for the sort of data mining that a neural network does, then we are logically committed to the thesis that SPSS does semantic cognition when we tell it to factoranalyze a data set. For anybody remotely familiar with either SPSS or factor analysis, this would seem a position sufficiently absurd to dismiss out of hand.