When a neuroscientist claims that a population of neurons in macaque area MT represents the direction of visual motion, or that hippocampal place cells represent spatial location, something philosophically peculiar is occurring. A concept borrowed from folk psychology—the idea that mental states are about things—has migrated into the technical vocabulary of a mature empirical science. Whether this migration is legitimate remains one of the deepest unresolved questions in the philosophy of cognitive science.

The stakes are considerable. If representational talk is merely heuristic scaffolding—a convenient shorthand that will eventually be discarded like phlogiston or the four humors—then contemporary computational neuroscience rests on a shaky conceptual foundation. If instead representation names a genuine natural kind, we must specify precisely what neural feature distinguishes a state that refers to something from one that merely covaries with it.

The challenge is sharpened by dynamical systems theorists, ecological psychologists, and radical enactivists who argue that representation is explanatorily dispensable. Neural systems, on their view, are better modeled as coupled dynamical processes than as manipulators of internal symbols. Evaluating this claim requires careful attention to actual explanatory practice in neuroscience—not merely armchair semantics. What follows examines the criteria proposed for genuine representation, the anti-representationalist critique, and whether representational explanation performs indispensable work in our best neurocomputational theories.

The core difficulty is distinguishing genuine representation from mere causal covariation. A thermostat's bimetallic strip covaries reliably with ambient temperature, yet few theorists want to say it represents temperature in any philosophically robust sense. Neurons in V1 covary with oriented edges in the visual field, but the same problem arises: what elevates this covariation to aboutness?

Dretske's informational semantics proposed that representation requires indicator functions—a state represents X when its biological function is to carry information about X. This teleosemantic move grounds content in selection history, whether evolutionary or developmental. Millikan's proper functions and Neander's etiological accounts extend this framework, tying representational content to what neural mechanisms were selected to do.

But teleosemantics faces the indeterminacy problem: for any neural state, multiple candidate contents typically satisfy the selection criteria. Does a frog's tongue-strike detector represent flies, small dark moving things, or nutritious targets? Fodor's asymmetric dependence theory tried to resolve this by appealing to nomological relations between error-tokens and veridical tokens, but critics argue the account presupposes the very semantic notions it aims to naturalize.

Structural resemblance theories offer an alternative. Cummins, Ramsey, and more recently Shea have argued that neural representations work by preserving structural relations between representing vehicles and represented domains. Grid cells in entorhinal cortex arguably instantiate this: their firing patterns preserve metric properties of physical space through a hexagonal tiling structure.

Yet each proposed criterion admits counterexamples, and no consensus has emerged. This persistent failure to naturalize content is either a temporary embarrassment awaiting theoretical progress or, more troublingly, evidence that representation is not a natural kind at all.

Takeaway

The failure to specify what makes a neural state genuinely about something—rather than merely correlated with it—may reveal that representation is a projected category rather than a discovered feature of nervous systems.

A vigorous tradition rejects the representationalist framework entirely. Van Gelder's dynamical hypothesis, elaborated by Chemero, Beer, and Thelen, holds that cognitive systems are better characterized as coupled dynamical systems governed by differential equations than as computational manipulators of internal symbols. The Watt centrifugal governor, not the Turing machine, becomes the paradigm.

Radical enactivists like Hutto and Myin push further, arguing that content-involving cognition is a late evolutionary and cultural achievement, largely absent from the basic sensorimotor coupling that constitutes most neural activity. On this view, neuroscientists who describe V4 neurons as representing color are committing a category error—importing folk-psychological vocabulary into a domain where it doesn't belong.

The ecological tradition inherited from Gibson emphasizes that organisms detect affordances through direct perception of environmental invariants, bypassing internal representations entirely. Modern proponents point to embodied and situated cognition research showing that many apparently cognitive tasks are accomplished through body-environment interactions rather than internal symbol manipulation.

These challenges carry genuine force. Reaching movements, locomotion, and many perceptual constancies can be modeled without positing representational intermediaries. Kelso's coordination dynamics and Warren's ecological approach to navigation demonstrate that rich behavioral competencies emerge from dynamical coupling without internal maps or symbols.

Yet critics like Clark and Shapiro argue that the anti-representationalist program overreaches. Offline cognition—planning, imagination, counterfactual reasoning—seems to require decouplable internal states that stand in for absent affairs. The dispute has increasingly narrowed to whether some cognitive phenomena genuinely require representational explanation, or whether sufficiently sophisticated dynamical models can accommodate all cognitive complexity.

Takeaway

The dynamical and enactivist critiques force us to distinguish which cognitive capacities genuinely require internal stand-ins for absent states of affairs, and which can be explained by real-time coupling alone.

The decisive question is pragmatic: does representational explanation do genuine scientific work, or could it be paraphrased away without explanatory loss? The evidence from contemporary computational neuroscience strongly suggests the former, though not without complications.

Consider the neural coding of decision variables in lateral intraparietal cortex. Shadlen and colleagues' drift-diffusion models describe LIP neurons as accumulating evidence toward decision thresholds. The predictive and explanatory power of these models depends essentially on treating firing rates as representing evidence for competing hypotheses. Purely dynamical redescriptions capture the trajectories but lose the crucial semantic connection to the perceptual discriminations being made.

Similarly, Bayesian brain frameworks—prominent in predictive coding accounts developed by Friston, Rao, and Ballard—treat neural populations as encoding probability distributions over hidden environmental causes. Here representation is not decorative: the computational role of prediction errors requires that they be errors about something, specifically the discrepancy between predicted and actual sensory input.

Deep learning has complicated the picture. Artificial neural networks trained on visual tasks develop internal states that resemble those found in ventral stream neurons, yet whether these states genuinely represent anything—or merely instantiate useful covariances—remains contested. The success of unsupervised representation learning suggests that representational structure can emerge without explicit teleological grounding.

The most defensible position may be a pluralistic realism: representation is a genuine but restricted phenomenon, applying paradigmatically to systems that exhibit decouplability, structural correspondence, and functional exploitation of that correspondence. Not every neural state represents, but denying representation across the board impoverishes our best explanatory frameworks.

Takeaway

Representational explanation earns its keep when neural states are exploited by downstream mechanisms precisely because of what they carry information about, not merely because of their causal profile.

The concept of representation occupies a contested but seemingly ineliminable position in contemporary neuroscience. Attempts to fully naturalize semantic content have not succeeded, yet attempts to eliminate representational talk from our best neurocomputational theories have proven equally unworkable for offline cognition and predictive processing.

The philosophically honest position recognizes that representation is neither a discredited folk-psychological holdover nor an unproblematic natural kind. It is a theoretical posit whose legitimacy varies across explanatory contexts. Where neural states are decoupled from immediate stimulation, exploited downstream for their informational content, and structured to preserve relations in their target domains, representational explanation earns its place.

This scientific-philosophical dialectic illustrates something important about naturalized epistemology: philosophical concepts are refined, not merely applied, through engagement with mature science. The neuroscience of representation is simultaneously advancing our understanding of the brain and reshaping what we mean by aboutness itself.