Concepts are the constituents of thoughts. As understood by cognitive scientists, they are items in semantic memory that are deployed in higher cognitive processes including inference, categorization, and judgment. In Chapter 2 of his excellent book Cognitive Ontology, Khalidi examines whether concepts form a unified kind, whether they are based in perceptual capabilities or amodal ones, and whether they can be individuated solely by factors internal to the agent’s mind. He adopts a form of pluralism that splits concepts into three kinds: prototypes (representations of statically-prevalent features of a category), theories (bodies of knowledge that serve as explanatory schemas), and modal processes (perceptual, motor, and affective simulations deployed in higher cognition).
Khalidi individuates concepts using his “functional” (56) or “causal-etiological” (72) account of concepts. He argues that theories, prototypes, and modal processes differ in that they: 1) occupy different levels in Marr’s hierarchy of analysis (computational, algorithmic, and implementational, respectively), 2) explain different phenomena (e.g., prototypes explain rapid judgments about category membership, while theories explain the criteria for possessing a concept), and 3) have different etiologies and causal profiles.
I found Khalidi’s use of etiology to individuate concepts in Chapter 2 particularly thought-provoking. Many theorists individuate cognitive kinds based on whether they correspond to distinct neural mechanisms or whether they are distinct kinds of representations. Khalidi argues that etiology and external factors matter as well.
He claims that children are often held to possess a concept such as alive based on the discriminatory abilities granted by prototypes (e.g., recognition of self-directed or animate motion). Adults are said to possess the same concept, but theirs is embodied in a richer theory (e.g., adults know that living things are not dead whereas children identify living things prior to their knowledge of death). Khalidi’s explanation of why psychologists attribute the concept alive to children, rather than a distinct and more basic concept, say alive*, is that: 1) children can communicate with adults about living things with their prototype-based concept, and 2) children are on a developmental trajectory that will result in them acquiring the adult concept. Thus, what makes a child’s prototype of alive a genuine concept is not the prototype itself qua mental entity but also externalistic factors such as embeddedness in a social-linguistic community.
Khalidi’s causal-etiological approach, which is one of the book’s central theoretical contributions, provides a fresh lens through which to view questions about cognitive ontology. It might, for example, provide new resources for forestalling eliminativism about concepts. Eliminativists (e.g., Machery 2009) argue that “concept” is not a fruitful member of our cognitive ontology since it refers to distinct representational kinds (e.g., prototypes, exemplars, and theories) that are deployed for different purposes and can yield conflicting judgments about the same semantic items. Theorists have often struggled with why mental representations with such differing properties should be united under one kind. Khalidi’s proposal that external factors may be constitutive of cognitive kinds may furnish new reasons to prefer pluralism over eliminativism. His causal-etiological approach could be productively applied to other debates about cognitive ontology, such as whether distinct memory types (e.g., episodic, semantic, implicit) form a legitimate higher-level kind.
Chapter 2 also contains an insightful review and critique of the modal theory of concepts in cognitive neuroscience. The modal theory claims that concepts consist of perceptual, motor, and affective simulations carried out in modal regions of the brain (e.g., visual cortex). While I sympathize with many of Khalidi’s points here—in particular, his concern that activation in modal areas in functional magnetic resonance imaging (fMRI) studies may reflect activity other than semantic processing—this section raises some concerns I have about how Khalidi formulates the relationship between neural structures and cognitive kinds.
First, I question his use of Marr’s levels to lump or split cognitive kinds. Modal theorists point to data such as fMRI activity in motor regions during verb processing as evidence that concepts are grounded in sensorimotor representations. Khalidi thinks this neuroscientific work occupies the implementation level while behavioral studies testing prototype theory occupy the algorithmic level. Thus, modal theorists are studying a different cognitive kind than behavioral psychologists. But I think for neural activity to truly count as conceptual, it would have to be describable in computational or algorithmic terms. For motor activity related to “kicking” to constitute one’s understanding of the verb “to kick,” that activity must have certain computational properties such as being evoked by visually dissimilar kicking episodes (e.g., it could be elicited by watching a horse or a human kick). I think any successful theory of concepts could be described at each level in Marr’s framework.
Second, I think Khalidi makes too much of apparent variability in the neural signatures associated with concepts. He denies that concepts can be analyzed in terms of neural implementation because there is too much variability in the regions activated during concept use. He similarly claims in Chapter 5 that episodic memory lacks a stable neural correlate and is generally opposed to identifying cognitive kinds with neural mechanisms (partly due to his externalist picture). But these inferences may go too quickly. Sometimes variability in brain activity associated with a cognitive kind occurs because researchers illicitly assume that different tasks measure the same cognitive construct (McCaffrey and Wright 2022). Furthermore, variability in the regions recruited during cognitive processing need not rule out stable neural correlates at some other level of analysis. Perhaps different concepts or memory episodes are indeed realized by different neural tissue, but that tissue may instantiate different tokens of the same neural type.
These qualms aside, I applaud Khalidi for providing resources to think about the cognitive ontology debate beyond representational kinds and neural mechanisms.
Machery, E. (2009). Doing without concepts. Oxford University Press.
McCaffrey, J., & Wright, J. (2022). 14 Neuroscience and Cognitive Ontology: A Case for Pluralism. In De Brigard and Sinnott-Armstrong (eds.), Neuroscience and philosophy. The MIT Press.
I’d like to respond to two points in Joe McCaffrey’s incisive comments. The first concerns how I view the difference between concepts proper, which are cognitive kinds, and their algorithmic and implementational counterparts, which I don’t think of as cognitive kinds, but kinds that pertain to different domains. McCaffrey writes that I claim “that children are often held to possess a concept such as alive based on the discriminatory abilities granted by prototypes (e.g., recognition of self-directed or animate motion). Adults are said to possess the same concept, but theirs is embodied in a richer theory (e.g., adults know that living things are not dead whereas children identify living things prior to their knowledge of death).” I would put it a bit differently. According to some developmental psychologists, children don’t acquire a theory-based (or computational) concept ALIVE until they are typically around ten years old (though six-year-olds can acquire it with training). However, both children and adults do seem to have a prototype that enables them to recognize the self-directed motion typical of living things. I argue that this latter is not the concept ALIVE, and that it involves a different psychological kind, call it the prototype. We can label them concept1 and concept2, as long as we keep in mind that they are genuinely different psychological kinds and shouldn’t both be thought of as different versions of the concept ALIVE. In this case, developmentalists tell us that children (under 10) don’t sufficiently differentiate between concepts like ALIVE and concepts like ANIMATE, ACTIVE, and REAL, and that is why we don’t ascribe the concept ALIVE to them. But in some cases, we do ascribe a concept to children before they are able to reason with it, answer questions satisfactorily, make appropriate inferences, and so on (e.g. ANIMAL). Why? I would express the first reason we do so differently than McCaffrey: children are in touch with the same external determinants as adults, and we individuate concepts partly based on their etiology. This is a standard externalist taxonomic practice. (I would put the second reason as McCaffrey states it.)
The second point is that McCaffrey argues that I make too much of apparent variability in the neural signatures associated with concepts. He thinks that “variability in the regions recruited during cognitive processing need not rule out stable neural correlates at some other level of analysis.” He adds that “sometimes variability in brain activity associated with a cognitive kind occurs because researchers illicitly assume that different tasks measure the same cognitive construct.” If I understand him correctly, McCaffrey is saying that when we find that there are different neural signatures associated with different tasks involving what is purportedly the same cognitive construct (e.g. the concept KICK), what may instead be going on is that there are different cognitive constructs at play. I’m not sure what these different cognitive constructs would be. Moreover, I think it’s pretty clear that in the particular case alluded to here (different tasks all deploying the concept KICK) there is a cognitive construct in common to the tasks, namely the concept KICK. This commonality is apparently not reflected at the neural level. So rather than say that there isn’t a single cognitive construct, despite the fact that all these tasks involve thinking about kicking, I think it would be more reasonable to conclude that the concept KICK doesn’t correspond to a neural kind.