1. Are objects really represented in multiple ways?

It is true that a variety of theoretical models (prototype, exemplar, theory-based) have been proposed to explain how concepts are represented and used. However, Machery's implication that each has equal explanatory power and therefore warrants status as a separate kind of concept is incorrect. While there are certainly limits to the things that exemplar models can explain (see Murphy Reference Murphy2002), there is also little doubt that exemplar models do a better job of explaining laboratory data on category learning (see Kruschke [Reference Kruschke, Lamberts and Goldstone2005] for a review) and other key phenomena (e.g., the effects of category-irrelevant features on classification, as in Allen & Brooks Reference Allen and Brooks1991) than models which only assume storage of prototypical features.

The “theory” approach was originally proposed to explain aspects of conceptual experience that lie outside the purview of “data-driven” approaches, including sensitivity to conceptual coherence and the causal basis of categories (Murphy & Medin Reference Murphy and Medin1985). However, there have been significant advances towards integration of data-driven and theory-based approaches. Heit's (Reference Heit1994a) integration model incorporates prior knowledge into an exemplar model to successfully predict patterns of category learning in knowledge-rich domains (Carmichael & Hayes Reference Carmichael and Hayes2001). In Rehder and Murphy's (Reference Rehder and Murphy2003) KRES model, background knowledge is represented in a connectionist network to explain how data-driven category learning is accelerated in the presence of knowledge and how ambiguous features are reinterpreted in the light of feedback. Similar efforts to integrate data-driven and knowledge-based approaches have met with considerable success in explaining category-based induction (e.g., Kemp & Tenenbaum Reference Kemp and Tenenbaum2009). Note that these are not “hybrid models” in the sense used by Machery. They do not assume separate “exemplar” and “knowledge” modules. Instead, prior knowledge is represented in a format similar to that used for learning new exemplars (e.g., the integration model instantiates prior knowledge as retrieval of relevant exemplars from memory; KRES does it via feature to feature associations and similarity between known and novel concepts). These models show that the conceptual knowledge associated with “theory” and “exemplar” approaches need not be thought of as independent and that their interaction can be accounted for within a unified theoretical framework.

2. Different representations for different tasks?

One of the key sources of evidence cited in support of the heterogeneity hypothesis is that different aspects of object knowledge are activated in the service of different goals or tasks. Classification learning, for example, focuses attention on the differences between categories, whereas inductive predictions rely on knowledge of within-category structure (Markman & Ross Reference Markman and Ross2003). The first thing to note is that this kind of flexibility is not limited to concepts. It is a ubiquitous feature of human cognition. For example, changes in retrieval conditions lead to important changes in the kinds of memories that are activated (Roediger Reference Roediger2008), but this does not mean that we should abandon the notion of object “memory.”

Second, it is important to recognise that empirical findings pointing to the “context dependency” of concepts can be oversold. A consistent empirical finding is that although context is important, a core conceptual representation of a given object is often retrieved when the object is used in a variety of conceptual tasks (Markman & Ross Reference Markman and Ross2003). For example, taxonomic features appear to be activated by default in a variety of tasks involving biological concepts, whereas the activation of causal relations is dependent on task goals and domain experience (Shafto et al. Reference Shafto, Coley and Baldwin2007).

Finally, and most critically, it is possible to develop models that have the flexibility to explain context dependent changes in concept retrieval – SUSTAIN (Love et al. Reference Love, Medin and Gureckis2004) is a good example. Such models achieve flexibility by incorporating well-established psychological processes, such as selective attention and discrimination learning, and the use of a range of similarity metrics. However, they do so while retaining an assumption of a conceptual “core” in object representation. As well as explaining data that Machery claims are problematic for standard models, these approaches may be extended to examine how conceptual knowledge influences performance in domains such as problem-solving (Markman & Ross Reference Markman and Ross2003) and recognition memory (Heit & Hayes Reference Heit, Hayes, Sloutsky, Love and McCrae2008), which have traditionally been divorced from the study of concepts.

3. What kind of evidence is needed to test the heterogeneity hypothesis?

Machery suggests that the heterogeneity hypothesis is supported to the extent that we can identify experimental or neuropsychological dissociations across different conceptual tasks. Although similar views are frequently espoused in the psychological literature (e.g., Ashby & Maddox Reference Ashby and Maddox2005), they should be treated with considerable caution. Deciding whether a given data pattern supports models that posit a single causal process (e.g., exemplar memory) or multiple, independent processes is a complex and tricky business. Patterns of single or double dissociations in categorization performance across tasks or patient populations can be produced by single-process models (e.g., Newell & Dunn Reference Newell and Dunn2008). Rather than simply assuming that patterns of dissociation point to a particular kind of cognitive architecture, future progress in mapping conceptual representations will need to apply rigorous analytical techniques such as state-trace analysis (Dunn Reference Dunn2008) and a careful comparison of well-articulated single- and multiple-process models.

The empirical work summarised by Machery suggests that our concept representations are complex and multifaceted. However, we argue that the balance of empirical and modelling work shows that the notion of “a concept” remains a useful heuristic in psychological explanation.