Study notes: What Is a Concept? A Survey of Theories for the Age of Artificial Minds

Content: what a concept represents, and how. Format: what kind of representational vehicle carries it. Origin: whether it is innate, learned or culturally universal. Dynamics: how it changes, combines and is deployed.

In the broad sense cognition is information processing, from retinal edge detection to mathematical reasoning. In the narrow sense it is conceptual processing: categorization, inference, and representations that combine into thoughts with truth conditions. Some disagreements, including Block’s with Quilty-Dunn, turn on where perceptual processing ends and conceptual thought begins.

• Classical. A set of necessary and sufficient conditions. Wittgenstein’s GAME resists definition: games share family resemblances, not a common set of conditions.
• Prototype (Rosch). A statistical best example, with new cases categorized by similarity to it. This explains graded typicality.
• Exemplar (Medin and Schaffer, formalized by Nosofsky). The remembered instances themselves, with new cases compared to the whole set. Both similarity theories fail to compose: the prototype of PET FISH is not built from the prototypes of PET and FISH.
• Theory-theory (Gopnik and Wellman, elaborated by Murphy and Medin). A node in an intuitive causal-explanatory theory, which is why deep causal features outweigh surface ones. It leaves open where the first theories come from, and it risks circularity: a theory needs concepts to be formulated in.
• Atomism (Fodor). An unstructured atom whose content is a law-like relation between mind and world. An atom has no parts to learn it from (the “doorknob problem”), which drove Fodor toward the view that all lexical concepts are innate and experience only triggers them.

Machery argues that “concept” is not a natural kind. The label covers at least three kinds of structure, prototypes, exemplars and theories, each deployed for a different task: quick categorization, causal reasoning, recall of an instance. The essay accepts this diagnosis of the impasse and rejects the elimination. Concepts can be characterized functionally, as whatever plays the concept role: the structures in long-term memory that are activated by perception or by other concepts, compose into structured thoughts, and guide inference and action. The introduction sketched this role with ACT-R’s memory store and spreading activation. A functional target is stable whether or not the category is a natural kind. It opens what the essay calls the format turn: asking how concepts represent, not only what.

Productivity: there is no longest thought, which requires a finite stock of primitives and rules for combining them. Systematicity: whoever can think “John loves Mary” can think “Mary loves John”, so thoughts share constituents that can be rearranged. Compositionality: the meaning of a complex thought is determined by the meanings of its parts and how they are combined.

The hypothesis is that mental representations are discrete symbols combined by syntactic rules. It is a claim about vehicles. It does not say that thought is conducted in a natural language. Atomism is a claim about content, and the two are logically independent: the vehicles could have language-like syntax while the primitives have internal structure, and atoms could be stored in a non-linguistic format.

Shepard and Metzler found that the time to judge whether two shapes are identical grows linearly with the angle between them, as if subjects rotate one shape mentally. Pylyshyn held that apparent imagery is description in a language of thought, and that the image-like phenomenology is epiphenomenal. For Barsalou, concepts are grounded in simulations, partial reactivations of sensorimotor states, so conceptual processing is perceptual processing in the absence of the stimulus. For Prinz, concepts are proxytypes, perceptual representations that stand in for categories, and the vehicles of concepts are always perceptual in origin. The challenge is composition: two images cannot be overlaid to give a conjunction. The essay takes this to suggest that image-like and language-like formats must coexist.

Block. Perception trades only in iconic representations. Cognition can use discursive ones, and the border between seeing and thinking is the border between the two formats. Perceptual content that enters cognition is wrapped in a discursive “cognitive envelope”.

Quilty-Dunn. His “perceptual pluralism” finds language-of-thought structure within perception. With his co-authors he lists six properties that cluster as a natural kind: discrete constituents, role-filler independence, predicate-argument structure, logical operators, inferential promiscuity, abstract conceptual content. All but logical operators, which remain an open question, show up in the perception of attended objects.

Evidence and replies. In visual working memory, an object’s color and orientation can be lost independently. A picture-like representation would lose them together. Block assigns the discursive properties to the envelope. Quilty-Dunn answers that the visual system itself computes over object representations, to keep coherence across eye movements and to track objects, so the structure serves perception.

They claim that analogy-making is the core of all conceptualization: every categorization, even calling something a chair, is an analogical leap from prior encounters. Concepts are then never static data structures. They are patterns of analogical activity that stabilize for a while and are modified by the next encounter. The essay adds that analogical mapping needs parts and relations to map between, so it presupposes something like the language-of-thought properties. It also needs the fluid, context-sensitive similarity judgments that image-like representations support. Analogy might therefore be the process that bridges iconic and discursive formats.

“Breakfast” is polysemous: an event or a kind of food, two related senses. “Bank” is a homonym with unrelated meanings. People judged equivocating arguments over polysemous words significantly more valid than those over homonyms. This suggests a single vehicle for the polysemous concept, one that takes part in composition and inference without a determinate referent. Quilty-Dunn concludes that content is epiphenomenal: the causal work is done by vehicle properties such as syntactic structure, format and computational role. The synthesis infers that interpretability research asking what a representation means may be looking in the wrong place, and that the mechanism of cognition may be best understood at the vehicle level.

Wierzbicka’s semantic primes are meanings lexicalized in every known language that cannot be decomposed further, such as SEE, FEEL, WANT, KNOW, DO, HAPPEN, GOOD, BAD. Every other meaning can in principle be paraphrased from them, so concepts have structure all the way down to the primes, against Fodor. Carey’s core knowledge systems are innate, domain-specific systems for objects, number, agency and geometry, with limited compositional power.

Wierzbicka asks what is linguistically universal and Carey what is developmentally foundational. Carey’s object mechanics has no obvious counterpart among the primes, and the evaluative primes match none of her domains. Agency appears in both. Both hold that the base is a structured repertoire of experiential and interactive primitives. Carey answers the doorknob problem with Quinean bootstrapping: new concepts are learned by restructuring the representational system, not by accumulating instances.

The bedrock includes experiential notions (SEE, HEAR, FEEL), agentive-cognitive notions (WANT, THINK, KNOW) and causal-dynamic notions (DO, HAPPEN, MOVE). If Wierzbicka is right, the universal foundation is a repertoire of experiential, agentive and causal primitives and not abstract logical machinery. The essay reads this as cross-linguistic evidence for two claims of its authors’ earlier articles: that phenomenal acquaintance is a distinctive epistemic anchor, and that recursive self-modeling needs agentive conceptual resources. Every known language has lexicalized both kinds of notion.

A concept is a generative model, a probabilistic model of how category instances are produced. Learning a concept is Bayesian inference over a structured hypothesis space. The essay reads this as a computational-level formalization of theory-theory. Its criticism is that these are toy models. Each addresses one task with a hand-designed hypothesis space and no perception, memory management, attention, real-time dynamics or action. The hypothesis space, where the theoretical work lies, is stipulated by the modeler. The models give a normative ideal at Marr’s computational level and say nothing about format, about retrieval, or about how the hypothesis space is built.

They are too high-dimensional and context-sensitive to be Rosch’s prototypes. They do not store individual instances, so they are not exemplars in the traditional sense. They are distributed and lack discrete symbolic structure at the level of single neurons, so they are not Language of Thought atoms. They are not analogical engines in any obvious sense.

They may be a new kind of concept. They may implement something like proxytypes, grounded in statistical co-occurrence instead of perceptual experience. Or interpretability tools may be uncovering structure like the language-of-thought properties, which is the possibility the essay finds most interesting. The caution: a sparse autoencoder assumes that the right decomposition has few active features, which is already a commitment about what concepts look like. The structure found may partly reflect the tool.

Thinking about democracy usually involves perceiving tokens: a word read, heard or vocalized internally. The tokens do not resemble the content, but they keep the processing in contact with something real. The contribution is meta-epistemic. It goes to the truth-preserving character of the processing, through coherence calibration and error signals, and not to the truth of any particular thought. The ambient perceptual embedding of a biological reasoner may add a further layer of calibration. So grounding matters “architecturally rather than semantically”. For a language model the question becomes whether its processing has this coherence. The concern is that without perceptual calibration, the truth-preserving character of its abstract reasoning “may be more fragile than it appears”.

• Whether grounding extends to abstract thought. A grounding story for DEMOCRACY may give the concept’s causal history and not its content (synthesis).
• Whether the original grounding of proxytypes requires phenomenal acquaintance at some point. Prinz’s framework leaves this open, and the essay’s suspicion that it does is “a further commitment” (Prinz).
• Whether the trade-off between compositional rigor and analogical fluidity is necessary. Language models seem to have the fluidity and cognitive architectures the rigor, and this may be an artifact of current designs (synthesis).
• What lies at the base, where Wierzbicka, Carey, Fodor and Hofstadter give four answers. Language models start from tokens and cognitive architectures from designed primitives. Whether it matters that neither starts from experientially and agentively saturated primes is “perhaps the deepest question connecting concept theory to AI” (synthesis).