Philosophy of Cognitive Science

What it asks

Cognitive science was built on an analogy that turned out to be enormously productive: the mind is to the brain roughly as a program is to the machine that runs it. Thinking is the manipulation of internal representations according to rules, and the rules can be studied without knowing the hardware.

That idea unified psychology, linguistics, computer science, neuroscience and philosophy into one field, and it has been under sustained pressure for thirty years. The pressure comes from two directions: whether the representations are needed, and whether the boundary of the thinking system is the skull.

The recurring problems

Whether computation is enough. Searle's Chinese Room argues that manipulating symbols by rule never amounts to understanding them, however good the output. The replies are numerous and none is universally accepted; the argument's persistence is itself evidence that something is unresolved.

The frame problem. A system that reasons about the effects of an action must somehow ignore the overwhelming majority of facts that are irrelevant, without checking them to find out. It looks like an engineering difficulty and is a deep problem about relevance.

Where the mind stops. Clark and Chalmers argued in 1998 that a notebook a person relies on constantly plays the same role a memory does, and should count as part of the cognitive system. Accept it and the boundary of the mind is not the body; reject it and you must say what makes the difference besides location.

What representations are for. Embodied and enactive approaches argue that much intelligent behaviour needs no internal model at all, only tight coupling between a body and its environment.

How it developed

The field was assembled in the 1950s from computing, information theory, linguistics and psychology, on the shared assumption that cognition is information processing.

Fodor's language of thought hypothesis in 1975 gave the assumption a specific form: thinking happens in an internal symbolic system with a syntax, which is why thoughts are systematic and productive.

Connectionism challenged it in the 1980s with networks that learn without explicit rules or symbols. The dispute was fierce, appeared to subside, and has been decisively revived by machine learning, where systems that were supposed to be unable to handle language do handle it.

From the 1990s the embodied, embedded, enactive and extended approaches attacked the framework from the other side, arguing that treating cognition as internal computation ignores the body and the environment that make it possible.

Predictive processing is the most recent unifying proposal: the brain as a hierarchy of predictions corrected by error, which claims to cover perception, action and learning in one framework.

The positions within it

Classical computationalism, connectionism, embodied and enactive cognition, the extended mind, predictive processing. All are in the Schools section.

Where it stands now

Large machine learning systems have made the old dispute empirical in a way nobody expected. Systems with no symbolic architecture do things the symbolic camp argued required one, which is evidence about engineering and contested evidence about minds.

The philosophically sharpest current question is what it would take to establish that any such system understands anything - which is the Chinese Room again, now with something to point at.