Concepts at the Interface

In Concepts at the Interface, Nick Shea integrates a wide range of empirical and theoretical literatures in psychology, neuroscience, and machine learning to offer a bold, systematic, and largely plausible account of concepts. He argues that concepts are “labels” that can be freely combined within working memory to conduct “content-general” reasoning, yet they also function as “plug-and-play devices” that interface with widely distributed bodies of sensorimotor and affective information to allow transitioning between content-general inferences and a variety of “content-specific” processes such as sensorimotor simulations. Thus, concepts bridge different cognitive processes.Many Western and non-Western philosophers posit concepts as entities whose composition and manipulation helps explain thought, language, categorization, and inference. Most major philosophical accounts—from Plato and Aristotle through Descartes and Kant, and into mainstream analytic philosophy (including Frege, Russell, and Carnap)—tie concepts to judgment and inference, language and meaning, rational cognition, and propositional knowledge. In this tradition, concepts are the building blocks of discursive thought. They serve as the elements manipulated in logical reasoning. The paradigm thinker is the discursive, reflective agent, not an embodied or sensorimotor one.In contrast, other forms of cognition—including imagination, sensorimotor simulation, and affect—have often been considered nonconceptual, undertheorized, treated as “lower” cognition peripheral to “core” conceptual thinking, and left outside the philosophical mainstream. There are exceptions (e.g., Aristotle’s phantasia, or Hume’s impressions versus ideas), but often these still don’t treat imagery, motor control, or affect as central to the function of concepts. Some contemporary philosophers and psychologists have offered an alternative: they argue that non-discursive forms of cognition are primary and concepts emerge from or are grounded in sensorimotor interaction (e.g., Lakoff and Johnson 1980; Clark 1996; Barsalou 1999; Prinz 2002). Still, these are often presented not as a complement but as a replacement for the long-standing tradition in which concepts are proprietary tools for discursive, logical thought.Shea aims at reconciling these two strands of thinking. Concepts, for Shea, are mental labels, activated within working memory, that connect to stored bodies of information. These bodies of information are likely to be distributed across several more or less dedicated sensorimotor or affective systems. Concepts (mental labels) enable deliberate, conscious control over the way the information in distributed systems is combined and manipulated. For example, when deciding what to wear in the morning, one may use the concepts SHIRT and SHORTS to call up a wealth of sensorimotor information and play around with various possible color and material combinations before settling on one’s outfit for the day. In addition, the labels themselves, independently of their connection to sensorimotor information, can be processed using logical operations. If I will wear both a shirt and shorts, it follows that I will wear a shirt—I needn’t know the color to work that out.One of Shea’s contributions toward this reconciliation project is to bring out and develop the distinction between “content-specific” and “content-general” computation. Cognitively functional computations tend to “respect” the contents of the representations the computations operate over. Roughly, this means that true inputs will reliably yield true outputs (at least, reliably enough to be useful for the organism). Shea’s content-general computations are those that respect contents using “broadly-logical” operations; If P is true and Q is true, then P Shea provides an example in which one is disposed to transition from the concept DOG to the concept BARKS. This distinction runs through the book and helps isolate precisely one of its central puzzles: How can concepts mediate between content-specific and content-general computations?One achievement of Shea’s book is to catalog, precisely spell out, and attempt to solve, using clear and convincing distinctions, many considerations and complications that such an ambitious project must meet. Shea also incorporates a substantial body of empirical research that any comprehensive theory of concepts should engage with.Another achievement of the book is to move beyond existing paradigms that focus on so-called “online” thinking. Online thinking occurs when one is actively interacting with the external environment—for example, when performing a behavioral task. Shea develops a model of concepts that foregrounds their role in “offline” thinking—when one is in the armchair reflecting, planning, deciding, deliberating, philosophizing, and so on. To this end, he highlights the way concepts combine in the “cognitive playground” of working memory to produce novel representations, such as simulations of possible future or counterfactual scenarios.The existing emphasis on online thinking makes sense in the context of gathering behavioral data for psychological theorizing. Experimenters typically require an operationalized measure of behavioral success or failure that allows hypotheses about the underlying psychological mechanisms to be empirically tested. In this context, we may read Shea as suggesting that the wealth of empirical data obtained through such tests now places us in a position to begin to generalize from behavioral data to offline capacities. This is a welcome shift of emphasis, which encourages serious thinking about the role of concepts in otherwise overlooked offline capacities, such as pure imaginative play.A proper assessment of a theory of concepts must fit it within the debate about cognitive architecture. There are three main approaches: the classical language of thought hypothesis (CLOTH), connectionism, and neurocomputationalism (Colombo and Piccinini 2023). Roughly, CLOTH says that (at least some important) cognitive processes are digital computations over sentences in a formal language, connectionism says that cognitive processes are similar to activation patterns within artificial neural networks, and neurocomputationalism says that cognitive processes are neural computations over neural representations, to be understood by studying real neurobiological systems.Much of what Shea says aligns him with a neurocomputational approach. He often appeals to evidence from cognitive neuroscience, and he deploys neurocomputational models, such as multidimensional state-space models of sensorimotor representation, clearly and effectively. At other times, Shea sounds perilously close to CLOTH. He periodically appeals to a language of thought and writes, for example, “In reasoning we move from some conceptually-structured thoughts to others using a domain-general process rather like theorem proving in logic” (117). That reasoning is like theorem proving is a typical CLOTH tenet: “It would not be unreasonable to describe Classical Cognitive Science as an extended attempt to apply the methods of proof theory to the modeling of thought (and similarly, of whatever other mental processes are plausibly viewed as involving inferences; preeminently learning and perception)” (Fodor and Pylyshyn 1988: 20–21). When Shea sounds like a CLOTH theorist, he falls into the habit—common among CLOTH theorists—of providing no compelling evidence for the processes they hypothesize (e.g., that discursive inference is like logical theorem proving) and few, if any, details about how such hypothetical processes might be reconciled with neuroscientific evidence.Perhaps Shea is sympathetic to the language of thought hypothesis but is not convinced by its classical version. Here, it helps to distinguish between classical and nonclassical LOT hypotheses (Piccinini 2025). Classical LOT hypotheses require that language-like representations be digitally encoded so that digital computations can be performed over them. Digitally encoded representations are made of discrete states organized in a clearly ordered way such that position within a sequence makes a functional difference to the system’s cognitive capacities. This digital encoding of information is the core of the analogy between cognition and digital computation, which is the heart of CLOTH. As Gualtiero Piccinini (2025) spells out, however, there is compelling evidence that brains do not encode information digitally and that many other assumptions required by CLOTH are empirically unsound. In contrast, a nonclassical LOT hypothesis maintains that some neural representations are language-like in the sense that they encode some of the structure of natural language. This might be done by neural manifolds that encode hierarchies of features at different scales, which is how neural representations generally seem to work. While LOT is still in play, the evidence supports a nonclassical LOT that requires no formal language and no digital encoding.At points Shea appears to endorse nonclassical LOT. He writes that some representations “can enter into analogue computations” (41). He adds that positing “step-by-step classical computations” (168) leads us to an intractable “frame problem” in which the system must decide which stored information is task-relevant, meaning that we should abandon such computational posits. He discusses computational processes involving multidimensional state-space activation and graded similarity ratings. But this is all about special-purpose uses of concepts. Here, Shea is neurocomputationally perspicacious. In the context of general-purpose uses of concepts, Shea provides less neurocomputational insight. Here, his appeal to LOT sounds more classical, like logical operations over sentences in a formal language. To evaluate Shea’s model against neuroscientific evidence, it would be helpful to know whether Shea takes general-purpose computation to have a classical structure.The distinction between classical and nonclassical LOT can shed light on Shea’s characterization of concepts as labels. Shea distinguishes his view from theorists who endorse a “pointer” theory of concepts. As he points out, there are different versions of this pointer view, some classical (e.g., Gallistel and King 2009) and some nonclassical (e.g., Eliasmith 2013). On the classical pointer view, a concept is a digitally encoded variable that the system uses as an address of a memory register to access information stored in that register. Shea’s view differs in that pointers do not themselves encode information about the concept’s referents, whereas labels can. However, absent a neurocomputational account of what makes the difference, it is hard to know how to evaluate these competing claims. What evidence would bear on whether concepts encode information directly? Are labels digitally encoded words of a formal language? A little more computational detail would make the difference more apparent and facilitate testing Shea’s account.Additional computational detail would also clarify how one individuates labels. According to one work cited by Shea to illustrate his notion of labels, working-memory neural representations used by participants to retrieve and manipulate sensorimotor information at the beginning of an experimental trial “are not well correlated with representations later in the trial” (Bouchacourt and Buschman 2019). In other words, multiple different representations appear to be used during a trial. If working-memory neural representations change during a single trial, how are we to individuate labels? This looks like a pressing question for Shea, since he individuates concepts with respect to the tokening of the same label during an episode of thinking. We need some more detail about how labels are implemented neurocomputationally in order to understand precisely what counts as activating the same label.A nonclassical LOT approach may very well be just what Shea needs to clarify his notion of a label. Labels might be amodal, neural encodings of (disambiguated) natural-language words. There is plenty of evidence that human brains encode language and that each word has its own neural representation. Since natural-language words can be ambiguous, brains must be able to disambiguate them internally. There is also evidence that neural representations of language need not be tied to any sensory modality—they can be amodal (Fedorenko et al. 2024). Since brains can process natural language and natural language requires combining and recombining words, there is independent evidence that neural representations of words can be recombined as freely as natural language requires, which is all that any theorist of concepts should need. There is also independent evidence that neural representations of words are connected with distributed sensorimotor representations (e.g., Calzavarini 2025). This is the sort of connection that, according to Shea, allows concepts to function as an interface between discursive and nondiscursive cognitive processes. None of this requires that neural representations of words be digitally encoded, let alone encoded by using a formal language, as CLOTH hypothesizes. Thus, amodal neural representations of natural-language words appear to be the perfect candidate to ground Shea’s metaphorical labels within the neuroscientific mainstream.In sum, Shea’s book brims with insight and innovation. He expands the debate on concepts to the interface between discursive and nondiscursive cognition, how concepts can mediate between both, and how they can be deployed within offline imaginative and counterfactual thought processes. He expands both the range of problems we should apply concepts to and the range of evidence we should consider. His excellent book is required reading for anyone in the field. With a little more detail on neural representation to flesh out his notion of a label, his account can lead us to a deeper understanding of concepts.