You see a photograph of a cup and know what it is. You can also recognize it from its name or by feeling its handle in the dark. Each time, you receive different information. How do such different clues lead you to recognize the same object?
What we know when we recognize a cup
That recognition draws on what we already know, even when the incoming information changes. Timothy T. Rogers reviews a long argument about semantic knowledge: what we know about the meanings of things rather than what happened on one particular morning. For much of the late twentieth century, the mind was pictured as a collection of partly independent stores, with a special semantic system separate from perception, action and much of language.
That picture did not appear from nowhere. Category research showed that things grouped together often share bundles of properties, not one defining essence. Chairs differ, yet many have seats, backs and uses that overlap. Other evidence seemed to point toward domain-specific systems: some neurological patients had greater trouble with living things than artifacts, or the reverse. A theory of meaning had to respect both the flexibility of everyday understanding and these apparently selective impairments.
Rogers's central claim is a change of scale. Instead of searching for one box that contains the meaning of 'cup', researchers increasingly ask how learned mappings across sensory, motor and linguistic representations make the object usable in many settings. The shift does not erase specialized brain contributions. It asks how specialization can arise within a connected system.
Why similarity alone was not enough
Early models often grouped objects by shared features. That captures an important part of learning: a child can recognize a new chair because it resembles familiar chairs. Yet the question 'similar in what respect?' has no answer independent of a task. A tennis ball and an orange may share shape; an orange and a lemon may share botanical and culinary relations. Which comparison matters changes with the question we ask.
A second pressure came from brain injury and imaging. Category-specific impairments seemed to suggest a special store for animals or tools. But categories differ in many ways at once: in shape, motion, use, visual experience and the kinds of knowledge people need. A selective deficit could arise because one connected system is damaged at a point especially important for that kind of information. The pattern alone does not prove that knowledge began in a sealed module.
Domain-specific theories sought to explain genuine patterns of impaired knowledge. Rogers returns to that problem in examining how meaning is organized. The task is to build a connected account that can explain those patterns too, including why damage sometimes affects broad meaning and sometimes affects particular categories more. A useful theory must specify mechanisms rather than merely rename the phenomena.
Learning relations instead of filing facts
Rogers revisits a neural-network model associated with David Rumelhart and colleagues. It did not store a list of explicit propositions as the central form of knowledge. Units representing objects and questions fed activity through weighted connections toward possible answers. Learning adjusted those weights. Over many examples, the network could capture regularities: related objects tend to share some properties, while the task determines which property is requested.
This was an explanatory move, not a claim that a computer network is a miniature human brain. A child does not learn from neatly labeled rows in a table. The model showed, in a controllable setting, how apparently structured knowledge might emerge from distributed learning rather than from a separate symbolic fact for every relation. Its successes and failures can then be compared with people's patterns of response.
A crucial issue remains: if different sensory and verbal systems learn separately, how do they converge on the same concept? The photograph, spoken word and physical action must be connected without pretending they are identical. That question motivates the next generation of models. It also makes the cup in the opening scene more than a metaphor: each route supplies different input to a network that must support coherent use.
A hub can connect different spokes
The hub-and-spoke approach gives that convergence a more explicit architecture. 'Spokes' involve modality-specific information such as how something looks, sounds, moves or is used. A transmodal 'hub' helps coordinate patterns across those spokes. Evidence from semantic dementia is important here: damage to anterior temporal regions can produce broad difficulties with meanings across categories and ways of testing, not only with one sensory channel.
Calling this region a hub does not mean it is a tiny dictionary with one entry per object. In the account Rogers reviews, its value lies in learning mappings across distributed features so that experience with one form can inform another. Loss of that coordination can undermine recognition, naming and other semantic tasks together. This contrasts with deficits that depend more strongly on the modality or task used.
The account also faces a hard question: why do some studies find different patterns for animals and artifacts? Rogers discusses unified models that preserve distributed learning while taking anatomical biases and differences in experience seriously. Brain responses to object categories can appear even when people hear names rather than see pictures; some patterns are also found in people born blind. Such findings make a purely visual explanation inadequate without forcing us back to separate sealed stores for every domain.
Knowing everything would be an unusable kind of knowledge
A network that connects everything it has learned still faces a practical problem. When you hear 'hammer', you do not immediately name it, mime its use, describe its shape and list every occasion you saw one. You select what the present task requires. Rogers therefore reconnects semantic representation with control: the processes that focus retrieval on useful relations and inhibit strong but irrelevant competitors.
Evidence from some people with left-hemisphere stroke makes this distinction visible. They can show broad semantic difficulty, but their performance may be especially poor when a task demands selection among competing meanings. They may respond inconsistently across tests and benefit markedly from a cue—or be misled by the wrong one. This pattern differs from the more stable degradation of semantic knowledge associated with semantic dementia. Both can affect meaning, but not in the same way.
The distinction explains a familiar experience without reducing it to brain jargon. If asked to name a picture of a hammer, you retrieve the word. If asked to use it, you activate movements and goals. If asked why it is dangerous in a particular room, you focus on a different relation. The object did not change; the route through knowledge did. Representation and control must be studied together.
How can scattered glimpses become one concept?
Control introduces a deeper puzzle about learning. The properties of a bird do not all appear together in one scene. A bird at a nest is not flying; a bird in flight is not laying an egg. If attention selects only what matters now, the learner cannot simply record every feature at once and notice their covariance. How does the system build an integrated concept from partial, task-dependent glimpses?
Rogers discusses work that gives learning and control a shared architecture. The central insight is that integration may depend on how representations are repeatedly used, not on a miraculous simultaneous presentation of every fact. Contexts and questions constrain what becomes active; learning across many such episodes can gradually link what no single episode contains. This is a proposed mechanism to investigate, not a description of one identifiable moment when a child 'gets' a concept.
It also changes how we think about mistakes. If a person brings the wrong aspect of a concept to a task, the problem may lie in selection, in a degraded representation or in the learned mapping between them. Similar wrong answers can therefore have different causes. The review's movement from modular stores to interaction is useful precisely because it creates more discriminating questions about behavior and brain injury.
Brain imaging and AI reopen the question
Recent imaging studies can examine meaning in richer situations than naming one picture at a time. People listen to stories or watch films while researchers compare changing brain activity with representations of words, scenes and events. Computational models make such comparisons possible, but an alignment between a model's vectors and brain signals is not, by itself, a complete explanation of understanding. It is evidence to interpret alongside behavior, development and clinical patterns.
Large language and vision models add another test bed. They can perform some tasks once used to probe human concepts. Rogers argues that matching human accuracy is not enough for cognitive theory; the revealing evidence often lies in patterns of similarity, error and generalization. When an AI system differs from people, that difference may show which relations human learning uses that the model lacks. When it resembles them, researchers still need to ask which process produced the resemblance.
Rogers brings together findings on learning, brain injury and computational models to explain how knowledge develops. Knowledge is learned in a connected system, shaped by sensory and linguistic experience, anatomical constraints and the control demands of the moment. The cup in the opening example is understandable across sight, touch and words because those paths can meet. Explaining exactly how they meet remains the research program.
What to keep
- Meaning emerges from learned relations across systems.
- Semantic representation and task-dependent control interact.
- AI offers models to test, not a settled account of human understanding.
The nuance
What this story does not let us claim
A distributed account does not mean every brain area does the same job, nor that today's AI models understand as humans do.
Go deeper: concepts and optional reading practice
Reversible cards
Test the concept
Topic assessment
Check whether you can explain the idea, not merely recognize it.
Every answer is grounded in the explanation above. Some questions introduce new cases to test whether you can apply and analyze the ideas, not merely recall them.
Source & scope
Understanding Understanding: Reconnecting Theories of Human Semantic Knowledge. Timothy T. Rogers. Annual Review of Psychology, vol. 78 (2027). First published as a Review in Advance on October 2, 2026.
This editorial guide is an independent explanation. Consult the original publication for the studies, references and full context.
Open the original review ↗