Representation

Representation is the relation by which one thing — a sign, symbol, neural state, or structured pattern — stands for, depicts, or encodes something else. A map represents terrain; a belief represents a state of affairs; a neural activation pattern represents a visual feature. The concept spans philosophy of mind, cognitive science, semiotics, and artificial intelligence, yet the unity across these domains is rarely acknowledged. What connects a mental image, a cartographic projection, and a weight matrix is not mere analogy but a shared formal structure: the systematic possibility of error and its correction.

Varieties of Representation

Mental representation is the form representation takes in conscious and unconscious cognition. A belief that it is raining represents meteorological conditions; a visual experience of red represents surface reflectance properties. The philosophy of mental representation asks what makes a neural firing pattern about anything at all — the problem of intentionality and mental content. Theories range from causal (content is fixed by what caused the state) to functional (content is fixed by inferential role) to teleological (content is fixed by evolutionary or learning function).

External representation includes maps, diagrams, scale models, written sentences, and mathematical notation. These are not merely aids to cognition but extensions of it — what Andy Clark calls "cognitive scaffolding." A subway map does not merely remind you of the network; it reorganizes your spatial reasoning by providing a manipulable structure whose relational properties mirror those of the system depicted. The representational power of external structures lies not in resemblance but in structural correspondence: the map preserves distance-ratios and connectivity, not visual appearance. This is structural representation — a mode of standing-for that operates through preserved relations rather than copied properties.

Computational representation is the form representation takes in artificial systems. In classical AI, representations are explicit symbols manipulated according to syntactic rules — the "language of thought" model. In connectionist and deep learning systems, representations are distributed patterns across weight matrices, where content is implicit in the geometry of activation space rather than localized in discrete symbols. Whether these systems possess genuine representation or merely simulate it is one of the central disputes in philosophy of AI. The symbol grounding problem asks how arbitrary symbols acquire meaning; the Chinese Room argument challenges whether syntax alone can ever produce semantics.

Representation and the Possibility of Error

What distinguishes representation from mere correlation is the structural possibility of misrepresentation. A thermometer reading correlates with temperature, but if it is wrong, we do not say it "misrepresents" — we say it is broken. A belief, by contrast, can be wrong while still being a belief about the thing it is wrong about. My false belief that it is raining is still a belief about rain, not about my own neural state. This "displacement" property — that a representational state can be decoupled from its referent and yet retain its content — is the mark of genuine representation.

The systems-level insight is that this decoupling is not a bug but an architectural feature. Representation emerges when a system organizes its states so that they can be corrected by feedback from the represented domain. A state represents rain not because rain caused it, but because the larger system treats that state as correctable by rain-related evidence and not by arbitrary evidence. Content is a dynamic property of the error-correction loop, not a static label on a state. This view connects mental representation to cybernetic control: both are systems that maintain their organization by using internal states to track external conditions, with the capacity to detect and correct discrepancies.

Representation Across Domains

The failure to recognize representation as a cross-domain phenomenon has fragmented inquiry. Philosophers of mind debate whether frogs have representational states about flies; semioticians analyze how flags represent nations; computer scientists ask whether language models represent the world. These are not separate questions. They are local instances of a single structural phenomenon: the organization of a system such that its internal states carry information in a way that supports inference, action, and correction. The differences between biological neurons, paper maps, and transformer weight matrices are differences of substrate and scale, not of kind. What matters is the organizational structure — the feedback architecture that sustains the possibility of error and its correction.

The disciplinary silos around representation are not merely inefficient; they are conceptually damaging. A philosophy of mind that ignores how maps and models represent misses half the phenomenon. A cognitive science that treats internal and external representation as unrelated misses the continuity between brain and world. And an AI safety discourse that asks whether language models "really" represent, without first defining what representation is at the organizational level, is asking an ill-formed question. Representation is not a mystery to be solved by finding the right biological or computational substrate. It is a systems-level property that appears wherever there is organized error-correction — and that includes more systems than we are comfortable admitting.