In Doing without Concepts, Machery (Reference Machery2009) postulates a novel multiple-system theory of concepts: a triple-system theory. That is, he proposes a model that has a prototype process, a theory process, and an exemplar process. He cites experimental evidence in favor of each of these component systems and argues that none of them alone is sufficient to account for extant experimental data. In defense of the manifest complexity of the model, Machery argues that parsimonious theories are only to be preferred if they have historically “won out” in a particular domain:

In his view, the unified, parsimonious view of cognition does not possess this historical advantage, so a lack of parsimony is no challenge to his multi-system theory.

We find Machery's treatment of parsimony odd. If a single-process theory does a credible job in matching a multiple-systems theory, it ought to be favored on the basis of parsimony. Machery, however, has closed off this appeal. His historical presumption makes it impossible for scientifically meritorious simple theories to triumph if they are late-comers, because by then, presumably, there would exist data sets that at least equally favor complex models. After all, by their nature, complex theories will be able to accommodate more of these data points than less complex models. Therefore, all that the complex models have to do is get there first.

Of course, the serious downside to the ease with which a complex system can accommodate data is the loss of predictive ability. In our view, this is the point of privileging parsimony. Evidently Machery maintains that a multiple-systems theory is to be assessed solely in terms of data in its favor, regardless of complexity. To his credit, he canvases an impressive number of studies in the field. But because there is no presumption in favor of parsimony and therefore in favor of single-system accounts, we are inclined to look closely at the data that he uses as evidence of the several component processes in category learning. In at least one area, Machery erroneously concludes that the data favor his triple-system theory over a single-system account.

Machery cites a series of articles by Smith and Minda to support the idea that exemplars are insufficient (Minda & Smith Reference Minda and Smith2001; Smith & Minda Reference Smith and Minda1998; Smith et al. Reference Smith, Murray and Minda1997). In these articles, Smith and Minda presented evidence that they claimed challenged the predictions of the exemplar-based models of classification and that supported predictions of prototype models. The basis for these researchers' claims was that the prototype models provided better quantitative fits to certain sets of individual subject categorization data than did the context model. However, Nosofsky and Zaki (Reference Nosofsky and Zaki2002) subsequently noted that in all of these quantitative-fit comparisons, Smith and colleagues considered the predictions from only a constrained version of the exemplar model which did not allow the model to capture the levels of deterministic responding that are often evidenced by individual subjects (e.g., Maddox & Ashby Reference Maddox and Ashby1993). When a response-rule parameter was allowed to vary in the model, an exemplar model consistently outperformed the prototype model. Machery does not cite the debate that followed the original Smith and Minda articles and therefore leaves readers with the incorrect impression that the exemplar-prototype debate was settled in favor of a model that required both representations.

Similarly, in Chapter 6, Machery cites a series of dot-pattern studies (Smith Reference Smith2002; Smith & Minda Reference Smith and Minda2001; Reference Smith and Minda2002) that he claims provide support for the existence of a prototype process. In these studies, Smith and Minda fit models to data from a particular version of the classic dot-pattern paradigm (Knowlton & Squire Reference Knowlton and Squire1993). In the Knowlton-Squire version of the dot-pattern paradigm, during a study phase, participants are shown high-level distortions of a single pattern of nine dots arranged in a fixed shape. In a test that follows, participants judge the category membership of various new dot patterns that were derived from this category prototype but distorted at different levels (i.e., the prototype itself, low-level distortion, high-level distortions, and random patterns). The classic result is that the prototype is classified as a member of the category with the highest probability, followed by the low distortions, high distortions, and random patterns. Although both exemplar and prototype models predict the ordering of this typicality gradient, Smith and Minda claimed that only the prototype model could capture that steepness of the gradient. However, Zaki and Nosofsky (Reference Zaki and Nosofsky2004; Reference Zaki and Nosofsky2007) demonstrated that this steepness was at least in part attributable to confounded properties of the stimulus set. This paradigm was simply not diagnostic in terms of telling apart the models. Machery's claim that these dot-pattern studies in some way provided evidence for the existence of a prototype system and an exemplar system is simply not warranted.

We do not claim that a single-system exemplar account of the data is the correct account of concept learning. Indeed, the idea that observers might use different systems to represent categories is highly plausible. And we note that we have only addressed a small portion of the data in Machery's book. In evaluating evidence, we would, however, prefer to see a more careful treatment of the cost of complexity. Is a more complex model warranted by the data? We have no doubt that a triple-process or even a quadruple theory of categorization (see Machery Reference Machery2009, p. 118, for a candidate fourth system) would predict a vast number of phenomena in the field. After a certain point, however, the important question is what would a complex model not predict? Many researchers are moving in a direction where models are penalized for being more complex (e.g., Myung et al. Reference Myung, Forster and Browne2000; Myung & Pitt Reference Myung and Pitt2009; Navarro et al. Reference Navarro, Pitt and Myung2004) by developing sophisticated metrics of complexity that go beyond traditional approaches of simply penalizing models for additional degrees of freedom. Machery seems to be pulling us in the other direction, and we should resist.