# Computationalism

> A family of theories that explains some or all cognition as computation implemented by a physical system.

- HTML version: https://robbiepalmer.me/ideas/computationalism
- Source: https://plato.stanford.edu/entries/computational-mind/

Computationalism is a family of theories that explains cognition as
computation. Its central question is whether mental processes can be described
by rules that transform representations or other information-bearing states.
Some versions apply only to particular capacities, such as vision or language.
Stronger versions claim that computation explains every important mental
process.

This is primarily a claim about minds. A theory of cognition could be
computational even if physics is not. A computational account of physics would
not by itself show that thinking is best explained in computational terms.

## The computational universe

The claim that reality is fundamentally computational usually appears under
the names **digital physics**, **digital ontology**, or **ontic
pancomputationalism**. These positions differ. Digital physics treats the
universe as a discrete computational structure. Ontic pancomputationalism says
that physical systems compute as part of what they are. The simulation
hypothesis makes another claim: our apparent universe is the output of a
computer in some other reality.

Konrad Zuse, Edward Fredkin, and Stephen Wolfram have defended versions of a
computational universe. David Deutsch belongs nearby, but his claim is more
precise. The Church-Turing-Deutsch principle connects what can be computed with
what physical systems can simulate. Deutsch used quantum theory to describe a
universal quantum computer. That makes computation a physical question. It
does not establish that the universe runs as one literal, clocked program.

One digital-physics proposal models the universe as discrete state transitions,
treats Planck time as a clock cycle, and explains time's arrow through
successive updates. Planck time is a natural unit derived from physical
constants. The clock and updates are additional hypotheses.

## What a computational explanation claims

A computational model identifies states, rules for moving between them, and a
physical system that implements those rules. The same abstract computation can
have different physical implementations. A calculation performed with silicon,
relays, or marks on paper can preserve the relevant organisation despite the
different materials.

The current case for computational intelligence looks different from the old
picture of hand-written rules, knowledge graphs, and Prolog programs. Large
language models use neural networks trained on statistical prediction. A
transformer learns distributed representations and weighted relationships from
data instead of receiving a complete symbolic theory of language from its
programmer. Statistical learning is still computation. It changes where the
structure comes from.

This matters because large language models can produce coherent language and
solve unfamiliar tasks without an explicit rule engine for each task. They
weaken the claim that systematic behaviour requires a hand-built symbolic
system. They do not by themselves prove that human cognition works like a
transformer. An engineered computational system can reproduce a capability
without reproducing the mechanism used by a person.

Classical theories remain part of the history and taxonomy. They treat thought
as rule-governed operations on structured representations. Connectionist
theories model cognition through networks whose changing weights and
activations produce a result. Neural computation ties a model more closely to
the organisation of a biological nervous system.

These programmes compete and overlap. Each defines computation differently.
Structural accounts focus on the causal organisation shared by
different implementations. Mechanistic accounts identify components and
operations that perform a task. Pluralists use different notions of computation
for different explanatory jobs.

## Computation needs an implementation

An abstract model does not explain a mind until there is an account of how a
physical system implements it. Without constraints, almost any object can be
mapped onto some computation. A wall contains many changing physical patterns,
but finding a mathematical correspondence between those changes and a state
table does not show that the wall runs a useful program.

Computationalists answer this triviality objection by adding requirements. The
physical states may need the right causal relations, dispositions under
possible inputs, representational function, or mechanistic organisation. The
choice matters. Different implementation criteria support different forms of
computationalism.

Meaning creates another problem. Formal rules can operate on a symbol because
of its shape or position, while thought concerns what the symbol represents.
An account that explains only formal transitions owes a story about how those
states acquire content. Some theories include representational content in the
computation. Others give computation a narrower role within a wider account of
perception, action, and environment.

## Limits and competing explanations

The strongest criticisms target the scope of computational explanation. Human
reasoning depends on relevance, context, and judgement about which of many
possible inferences deserves attention. A formal procedure can model a selected
task. That success does not establish that choosing and framing the task follow
the same procedure.

A brain-only account is too narrow. Cognition develops through a whole nervous
system in a living body, with perception and action forming continuous loops
through its environment. Embodied cognition studies those coupled dynamics
across psychology, neuroscience, robotics, and artificial intelligence. The
behaving organism becomes the explanatory target.

## A philosophical claim with scientific consequences

Computationalism itself is a philosophical interpretation. No single
experiment can apply a universal scientific method to it and return a verdict.
Philosophy of science helps define what counts as a model, an explanation, and
evidence in the first place. Requiring those standards to justify themselves
only by their own standards would be circular.

Specific computational models are still scientific claims. Researchers can
name a cognitive capacity, define the computation, state how a physical system
implements it, and compare its predictions with rival models. Philosophy is
judged more broadly through coherence, explanatory reach, compatibility with
scientific knowledge, and the research questions it makes possible. The levels
inform one another without collapsing into one method.

[Falsifiability](/ideas/falsifiability) applies most directly to the specific
models. [Theory-ladenness](/ideas/theory-ladenness) explains why the choice of
model and evidence already depends on a conceptual framework.
[Epistemology](/ideas/epistemology) asks what a successful model lets us know
about a mind rather than merely predict about its output.

## Sources

* Rescorla, Michael. ["The Computational Theory of
  Mind"](https://plato.stanford.edu/archives/spr2024/entries/computational-mind/).
  *The Stanford Encyclopedia of Philosophy*, 2024.
* Turing, Alan M. ["On Computable Numbers, with an Application to the
  Entscheidungsproblem"](https://doi.org/10.1112/plms/s2-42.1.230).
  *Proceedings of the London Mathematical Society* 42, 1936.
* Searle, John R. ["Minds, Brains, and
  Programs"](https://doi.org/10.1017/S0140525X00005756). *Behavioral and Brain
  Sciences* 3, no. 3, 1980.
* Piccinini, Gualtiero. *Physical Computation: A Mechanistic Account*. Oxford
  University Press, 2015.
* Piccinini, Gualtiero, and Corey Maley. ["Computation in Physical
  Systems"](https://plato.stanford.edu/entries/computation-physicalsystems/).
  *The Stanford Encyclopedia of Philosophy*, 2021.
* Deutsch, David. ["Quantum Theory, the Church-Turing Principle and the
  Universal Quantum
  Computer"](https://www.daviddeutsch.org.uk/wp-content/deutsch85.pdf).
  *Proceedings of the Royal Society A* 400, 1985.
* Vaswani, Ashish, et al. ["Attention Is All You
  Need"](https://arxiv.org/abs/1706.03762). *Advances in Neural Information
  Processing Systems* 30, 2017.
* Shapiro, Lawrence, and Shannon Spaulding. ["Embodied
  Cognition"](https://plato.stanford.edu/entries/embodied-cognition/). *The
  Stanford Encyclopedia of Philosophy*, 2021.
* Hossenfelder, Sabine. ["Theory and Phenomenology of Space-Time
  Defects"](https://doi.org/10.1155/2014/950672). *Advances in High Energy
  Physics*, 2014.

## Related ideas

- [Epistemology](https://robbiepalmer.me/ideas/epistemology.md): How we know what we claim to know, what justifies a belief, and when new evidence should change it.
- [Falsifiability](https://robbiepalmer.me/ideas/falsifiability.md): A claim is falsifiable when some possible observation could count against it, though real tests examine a bundle of assumptions.
- [Theory-ladenness](https://robbiepalmer.me/ideas/theory-ladenness.md): Observations depend on prior ideas about what to notice, how to measure it, and what the result means.

## Where it appears

- Blog post: [The Philosophy of Data Science](https://robbiepalmer.me/blog/2022-03-02-the-philosophy-of-data-science.md)

---

Markdown index of this site: https://robbiepalmer.me/llms.txt