The Computational Theory of Mind

The computational theory of mind explains cognition as rule-governed transformation of representations. Perception supplies representational inputs, reasoning transforms them, memory stores them, and action follows from outputs. The theory does not say minds are ordinary desktop computers. It says some cognitive processes are computational in a sense that abstracts from the material implementing them.

Computationalism earned its influence by explaining productivity, systematicity, and mechanistic decomposition. Finite systems can generate indefinitely many structured thoughts because operations are sensitive to their constituent form. Its difficulties concern content and relevance: formal rules operate on vehicles, while cognition is about a world and selects contextually important information from enormous possibility spaces.

Computation and cognition

A computation maps formally characterised states according to rules. The same abstract mapping can be implemented in electronic, mechanical, neural, or other physical media. This fits Functionalism: cognitive kinds can be identified by causal organisation rather than one material type.

Three claims should be separated:

1. Computational description: a process can usefully be modelled computationally.
2. Computational mechanism: physical components implement transformations over vehicles in virtue of their organisation.
3. Computational sufficiency: implementing the right computation is sufficient for possessing the relevant mind or mental state.

The first is weak and widespread; the third is a strong metaphysical thesis. Success of a model does not by itself establish sufficiency for understanding or consciousness.

Representations and vehicles

A computational state must be individuated in a way that supports transformation. Vehicles are the physical or functional states manipulated. Contents are what those vehicles represent. Syntax concerns formally identifiable structure; semantics concerns correctness and aboutness.

The distinction prevents a homunculus picture. A system need not read its symbols as a person reads text. Causal mechanisms respond to vehicle structure, while causal, teleological, inferential, and social relations fix content. The naturalisation problem is treated in Naturalizing Mental Content.

The challenge is semantic efficacy. If operations depend only on formal vehicle properties, how does what a representation means help cause a rationally appropriate result? A reply identifies content with or grounds it in the functional relations that the mechanism uses. External content remains harder because local computation can be shared by states with different wide meanings.

The language of thought

Fodor's language-of-thought hypothesis proposes an internal representational system, often called Mentalese, with combinatorial syntax and semantics. Propositional attitudes relate subjects to structured inner representations.

The hypothesis explains:

• productivity: finite primitives generate unbounded thoughts;
• systematicity: capacity for John loves Mary is connected to capacity for Mary loves John;
• compositionality: complex content depends on constituents and arrangement;
• causal precision: formal structure determines which transformations occur.

Public language need not be the medium of thought. Mentalese could be innate or learned and need not resemble English. The claim is that thought has constituent structure sufficient to support recombination.

Critics question whether imagery, motor skill, affect, and perceptual cognition share this format. A plural architecture may use language-like representation for some reasoning and distributed or analog formats elsewhere.

Classical architecture

A classical cognitive architecture stores explicit symbols and applies rules whose operation is sensitive to syntactic constituency. Processing may be serial or parallel; the defining point is that complex vehicles contain identifiable parts corresponding to semantic constituents.

This supports variable binding and general rules. From $Loves(x,y)$, a system can substitute different individuals while preserving relation structure. Planning and logical reasoning appear to need such compositional flexibility.

The architecture can be brittle. Explicit rules struggle with noisy perception, similarity, graceful degradation, and learning from examples. These weaknesses motivate Connectionism, whose representations are distributed across weighted networks.

Implementation

An abstract automaton does not become a computation merely because an observer maps its states onto symbols. A physical system implements computation through organised causal relations that support the relevant counterfactuals: had input differed, the system would have transitioned according to the rule.

This blocks trivial mappings under which a wall or weather system implements any finite computation by retrospective state assignment. Implementation requires principled state boundaries, causal transitions, temporal organisation, and often functional use of outputs.

The correct grain remains contested. At one level a processor performs arithmetic; at another it follows electronic dynamics; at another every physical trajectory can be described as state transition. Computational explanation is objective only when the chosen mapping supports intervention and explanatory generalisation.

The triviality problem

Putnam and Searle argued that sufficiently permissive implementation criteria make almost every physical system implement every computation. If true, computationalism cannot distinguish minds from rocks.

Counterfactual and causal-structural accounts respond that actual-state sequences are insufficient. A genuine implementation must preserve transitions under possible inputs, with components causally responsible for the mapping.

This response makes computation partly relational and functional. It also means a recorded lookup table and a flexible processor can produce the same actual outputs while implementing different counterfactual organisations. Behaviour on one run cannot settle architecture.

The frame problem

An intelligent system acting in a changing world must update what changed while retaining what did not. Explicitly listing every unaffected fact is computationally explosive. This is the classical frame problem.

In a broader sense, cognition must identify what is relevant among indefinitely many facts and possible consequences. Formal logic specifies valid entailments but not which ones matter now. A system can derive useless truths forever.

Proposals use default reasoning, non-monotonic logic, probabilistic models, learned attention, embodied skill, and task-specific heuristics. The problem pressures a purely rule-based picture but not computation as such; relevance-sensitive algorithms can still be computational.

The relevance problem

Relevance depends on goals, context, value, bodily needs, and social practice. These are not easily represented as a fixed database consulted before every action. Dreyfus and enactivists argue that skilled agents encounter a field already organised by significance rather than compute relevance from neutral facts.

Computationalists answer that goal states, learned value functions, predictive models, and resource-bounded search can produce selective processing. The relevant contrast is not formal calculation versus magic immediacy but explicit central search versus distributed, learned, and embodied computation.

The debate turns partly on explanatory vocabulary. Calling every adaptive dynamics a computation protects the theory at the cost of specificity.

Computation and consciousness

Computationalism about cognition does not entail computationalism about phenomenal consciousness. A system may transform representations, reason, and act while the hard problem remains.

A strong functionalist argues that the right computation is sufficient for experience. Absent-qualia and Chinese-room arguments challenge this. Biological and dynamical theories may require properties not preserved by abstract computational equivalence.

The distinction is essential for artificial minds. Showing that a machine implements a cognitive algorithm establishes a mechanism for a capacity, not automatically a subject who experiences it.

Levels of computational explanation

Marr distinguishes:

• computational level: what problem is solved and why;
• algorithmic level: which representations and procedures solve it;
• implementational level: how the procedure is physically realised.

The terminology is broader than computation in the first item: the computational level specifies an input-output task. The levels constrain one another. Hardware limits algorithms; representation shapes tractable problems; ecological goals explain why a capacity exists.

No level is automatically fundamental. A complete cognitive explanation may require all three plus personal-level reasons and social context. The later science folder will develop the general philosophy of levels; here the distinction prevents an algorithm from being mistaken for either a behavioural specification or a neural mechanism.

Assessment

The computational theory is best understood as a family of mechanistic hypotheses, not the claim that every mind is a stored-program computer. Its lasting insight is that formal organisation can explain cognitive capacities across materials. Its hardest task is showing how formal transitions become world-directed, context-sensitive thought.

Selected references

• Fodor, Jerry A. The Language of Thought (1975).
• Marr, David. Vision (1982).
• Newell, Allen and Herbert A. Simon. "Computer Science as Empirical Inquiry" (1976).
• Pylyshyn, Zenon. Computation and Cognition (1984).
• Searle, John R. "Is the Brain's Mind a Computer Program?" (1990).