Connectionism

Connectionism is the approach that models cognition with networks of many simple, interconnected units. Connectionists hold that information is stored not as symbols but in the strengths, or weights, of the connections between units, in contrast to the classical view that the mind is a symbol-manipulating computer (Garson & Buckner, 2019). The approach’s central dispute is with that classical view. Fodor and Pylyshyn argued that human thought is systematic, meaning that anyone who can think one thought can think structurally related ones, and that classical models explain this while connectionist models get it only by accident (Garson & Buckner, 2019). Critics argue that connectionism is therefore no viable alternative to classical computation (Rescorla, 2024), while a 2023 network matched human systematic generalisation in experiments (Lake & Baroni, 2023). The best-known empirical battleground is how people form the English past tense (Yoon, 2023).

History

Not covered by the research: early network models, the 1986 parallel distributed processing volumes, and the relation of connectionism to modern deep learning.

Core concepts

Networks and weights

In a connectionist network, what the network knows lies in its connection weights, not in explicit symbols (Garson & Buckner, 2019).

Classical computation

The classical computational theory of mind holds that the mind does not merely resemble a computer but literally is a computing system (Rescorla, 2024). Classical and connectionist views need not exclude each other: hybrid views hold that symbolic processing in the brain may be a matter of degree (Garson & Buckner, 2019).

Questions, controversies, and new developments

Can networks be systematic?

The Fodor–Pylyshyn argument holds that systematicity comes for free in classical models, because their symbols combine by rule, while in connectionist models it would be a lucky accident of training (Garson & Buckner, 2019). Marcus made a related empirical argument: the popular class of connectionist models that do without symbols cannot extend universal relationships to items outside the space they were trained on, so symbol manipulation may be needed (Marcus, 1998). On this basis, critics hold that connectionism does not furnish a viable alternative to classical computationalism (Rescorla, 2024).

In 2023, Lake and Baroni reported that a neural network trained specifically for compositional skills matched human systematic generalisation in behavioural experiments, challenging the claim that networks cannot be systematic (Lake & Baroni, 2023). It is a single study.

Rules or associations in the past tense?

The English past tense became a central test between rules and connections. In 2002 Pinker and Ullman, and McClelland and Patterson, set out the two positions (Yoon, 2023). Pinker and Ullman’s words-and-rules account holds that irregular forms are stored in declarative memory while regular forms are computed by a rule in the procedural system (Pinker & Ullman, 2002). The single-route alternative holds that one mechanism, analogy over stored examples, produces both (Blything et al., 2018).

The evidence divides:

• An ERP study found a brain response, left anterior negativity, to regular past-tense errors and phrase-structure violations but not to irregular errors, read as distinct processing of the two kinds (Newman et al., 2007).
• An N400 priming study found a regular-irregular dissociation, but argued that it fits single-system models better (Justus et al., 2008).
• In stroke patients, deficits with regular forms went with damage to regions linked to articulatory sequencing and phonological short-term memory rather than with grammatical difficulty (Meteyard et al., 2013).
• Children’s past tenses of novel verbs were explained by analogy over stored examples rather than by a formal rule (Blything et al., 2018).

Whether modern neural networks resolve the debate is an open question (Yoon, 2023). The Language article covers the same evidence from the side of language processing.

Replication record

Not covered by the research: replications of the key past-tense and generalisation findings.

Broader connections

Connectionism is one of the accounts compared in Architecture of the mind. Within predictive-processing theory, which is discussed under Embodied cognition, accounts range from cognitivist to strongly connectionist to dynamical (Allen & Friston, 2016). Connectionist networks share with associative learning the idea that knowledge lies in learned connection strengths.

Further reading

• Garson and Buckner (2019): the Stanford Encyclopedia of Philosophy entry, covering the systematicity debate and hybrid views.
• Rescorla (2024): the classical computational theory of mind and its critique of connectionism.
• Lake and Baroni (2023): a network that generalises systematically like people.