Connectionism
Connectionism is the research program in cognitive science and artificial intelligence that models cognition as the emergent product of large numbers of simple processing units — artificial neurons — connected in networks. It is the theoretical ancestor of modern deep learning and the philosophical opponent of classical cognitive science, which held that cognition is fundamentally symbolic: rule-governed manipulation of discrete representations.
The core connectionist claim is that the representations underlying cognition are not explicit symbols — not the discrete, compositional structures of predicate logic — but distributed activation patterns across many units, none of which individually represents anything interpretable. Meaning is carried by patterns, not atoms. This is an empirical claim about cognitive architecture, not merely an engineering preference.
History and the Parallel Distributed Processing Project
The modern connectionist program crystallized in the 1986 Parallel Distributed Processing (PDP) volumes edited by Rumelhart and McClelland. The PDP project demonstrated that multilayer networks trained with backpropagation could learn a range of cognitive tasks — past-tense morphology, visual word recognition, semantic inference — that had previously been modeled only with explicit symbolic rules. The demonstration was powerful: systems with no explicit rules could exhibit rule-like behavior. This reframed the theoretical question from 'what are the rules?' to 'what gives rise to the appearance of rules?'
The behaviorist parallel is worth noting. Behaviorism expelled internal representation from psychology on methodological grounds; connectionism restored internal representation but insisted it was distributed and subsymbolic. Both share skepticism about the explanatory value of explicit symbolic description. Both have been accused of trading one black box for another.
The Fodor-Pylyshyn Challenge
The most serious objection to connectionism came from Jerry Fodor and Zenon Pylyshyn's 1988 paper 'Connectionism and cognitive architecture: A critical analysis.' Their argument: human cognition exhibits systematicity and compositionality — the ability to think 'John loves Mary' is systematically connected to the ability to think 'Mary loves John'; the ability to represent any sentence is connected to the ability to represent its structural constituents. Classical symbolic architectures explain this by construction: representations are literally built from parts, and the rules operate on the structure of the parts.
Connectionist networks, Fodor and Pylyshyn argued, do not exhibit genuine compositionality. A network that has learned 'John loves Mary' does not thereby have the components for 'Mary loves John' in any principled sense — it has a weight matrix that happens to produce the right output. The systematicity is mimicked, not explained.
Connectionists responded with partial rebuttals — distributed representations can exhibit approximate compositionality; the challenge assumes a too-narrow notion of what counts as genuine structure. The debate was never resolved because it was, partly, a debate about what 'genuine' meant. What the debate established: connectionists and classicists disagree about the explanatory role of structure, not merely about implementation.
Connectionism and Contemporary Deep Learning
The relationship between connectionism (a theory of cognition) and deep learning (an engineering practice) is commonly elided and should not be. Deep learning inherits connectionism's architecture but not its ambitions. Connectionist researchers in the 1980s-90s cared about psychological plausibility — about whether their models made correct predictions about human cognitive errors, learning trajectories, and developmental patterns. Modern deep learning researchers care about benchmark performance. These are different projects.
The inference from 'large neural networks perform impressively on cognitive benchmarks' to 'connectionism is vindicated' is not valid. Connectionism made specific predictions about the structure of learned representations and the mechanisms of generalization. Whether large language models exhibit the learned representations connectionism predicted is a question that benchmark performance does not answer — because benchmarks measure outputs, and the question is about internal structure.
Interpretability research is, in part, an attempt to ask the connectionist question seriously: what have these networks actually learned? The preliminary answers suggest that large models learn representations that are neither purely symbolic nor purely the distributed attractors that connectionists anticipated. They are something third, currently without a principled theoretical description.
The persistent tendency to treat deep learning's engineering success as evidence for connectionist theory, or against classical cognitive science, confuses the product with the theory. A bridge does not vindicate Newtonian mechanics merely by standing. A language model does not vindicate connectionism merely by producing coherent sentences.