Three theoretical concerns regarding parallel distributed processing (PDP) modeling keep puzzling a logically minded cognitive scientist, if it is, as Rogers & McClelland (R&M) claim, designed to provide simulations of the acquisition and deterioration of human semantic cognition.
1. The weighting or strengthening of network relations in the base architecture (see Semantic Cognition, Rogers & McClelland Reference Rogers and McClelland2004, pp. 115, 356) is presumably a linear relation. It is hence transitive, that is, if x is heavier or stronger than y, and y heavier or stronger than z, then x must be heavier or stronger than z. But an arbitrary similarity relation does not need to be transitive, for a canary may be similar to an orange in color, and an orange similar to the moon in being round, but a canary is not therefore similar to the moon. If such similarity judgments constitute the basis for inductive generalizations and clustering (Semantic Cognition, p.183), do PDP networks require similarity to be strengthened to a transitive relation? Perhaps an item may be simultaneously included in two distinct clusters? To give a particular illustration of this problem: If the representation space in Figure 3.9 (Semantic Cognition, p. 112) were also to represent the color pink, how could it incorporate both salmon and rose in one shaded region?
2. According to Semantic Cognition (p. 182), basic names indicate the “labels that identify an object at an intermediate level of specificity” (e.g., bird), and general names “identify an item at a more inclusive level” (e.g., animal), and specific names “identify objects at a more specific level” (e.g., canary). So obvious logical entailment patterns should be validated; for example, if every bird is an animal and every canary is a bird, then obviously every canary must be an animal. Clearly, the list of attributes is highly structured by such logical entailments, which presumably facilitates children's review in allocating weights to attributes. Is there any evidence to assume that all the listed attributes are independently reviewed in assigning weights to connections, as suggested by linearly listing all attributes in PDP networks and requiring exhaustive review of these in attributing weights (e.g., the longer the list, the longer it takes to allocate weights)? Don't children quickly detect such valid logical patterns and use them in facilitating their review of attributes, skipping large sets of labels that don't apply simply because they depend on a label that is already known not to apply? In other words, doesn't logical knowledge help children in speeding up attribute weighting by making their review more efficient?
3. From an epistemological point of view, generic knowledge is special, “immunized” information. It is preserved in changing worlds and throughout a variety of contexts in virtue of a set of associated, generally accepted exceptions, that are prevented from serving as counterexamples falsifying the generically quantified statement (Carlson & Pelletier Reference Carlson and Pelletier1995). In learning to complete the similarly generic claim canary can…, any child will quickly exclude dead, frozen, or wind-up canaries in assigning a heavy weight to the attribute move. Having learned to treat a wind-up canary as an exception to the generic claim that canaries can move, the child applies the abstract notion of agency to anything capable of self-determined action (cf. Semantic Cognition, p. 131–33). If agency is considered to be one of Gelman's enabling constraints (Semantic Cognition, p. 133), then it seems undeniably to play a core role in concept formation at a pre-lingual and hence pre-explanatory phase of development. In the similarity-based learning algorithm of bootstrapping syntax (van Zaanen Reference van Zaanen2001; Reference van Zaanen2002), the child learns that if x moves y, then x causes y to move, and accordingly, y has been moved by x, and that therefore y has moved, but also that the reverse entailments do not hold. The causal source of the movement is consistently assigned as thematic role to the subject argument in either the intransitive or active transitive use of the verb move, as it is part of its lexical semantic content at least in ordinary English. As the child also knows that agency is required for self-movement, he or she would probably, at a linguistically adequate phase of development, be able to explain that a dead or wind-up canary, which he or she has observed to have moved from one location to another, must have been moved there by someone else. Such causal inference appears in no way logically prior to linguistic acquisition of the corresponding lexical items, nor does it seem to play a role similar to the logical relations alluded to above in structuring the set of attributes. But to conclude from such lexicalization of attributes that “Infants' sensitivity to a property, as indexed by attention, ease of learning, and other measures, is affected by the extent of its coherent covariation with other properties” (Semantic Cognition, p. 135), seems theoretically a bridge too far, although coherent covariation must be an extensional, observable counterpart to any obviously intensional logical entailment.
It is an interesting, and to my knowledge, novel claim that the PDP networks should also simulate semantic dementia effects. For instance, in overextending basic level names (dog), the model is claimed to be more likely to apply a highly familiar name incorrectly than a less familiar name to similar items (Semantic Cognition, pp. 215–17). It is well known that the familiarity of names is directly correlated with their early acquisition and hence retention in early memory, rather than with their frequency in recent usage (ter Meulen Reference ter Meulen, Reimer and Bezuidenhout2004). It would constitute an interesting new dimension in the empirical predictive power of PDP simulations, if time were an explicitly represented parameter. Consequently, the weighting should be an overtly dynamic relation, instead of a static time-slice model. In dynamic semantics a plethora of detailed representation systems have been developed that could well serve as inspiration for designing such overtly dynamic PDP systems (Kamp & Reyle Reference Kamp and Reyle1993; Lascarides & Asher Reference Lascarides and Asher1993; ter Meulen Reference ter Meulen1995; Reference ter Meulen, Higginbotham, Pianesi and Varzi2000; Reference ter Meulen, Zanuttini, Campos, Herburger and Portner2006; van Benthem & ter Meulen Reference van Benthem and ter Meulen1997).
