Connectionism
Connectionism is the approach to modeling mind and intelligence in which behavior emerges from the interactions of many simple units joined by weighted connections, with knowledge stored in the pattern of those weights rather than in explicit symbols and rules. It rose to prominence in the mid-1980s as a rival to symbolic AI, and its central bet, that learning in networks of simple units can explain what minds do, is the lineage from which today's deep learning descends.
The movement's founding document was the two-volume Parallel Distributed Processing: Explorations in the Microstructure of Cognition, published by MIT Press in 1986 and edited by David Rumelhart, James McClelland, and the PDP Research Group. Its chapters laid out a shared vocabulary of units, activations, connection strengths, learning rules, and parallel settling, and applied it to perception, memory, language, and cognitive development. It appeared alongside the Rumelhart, Hinton, and Williams backpropagation paper in Nature in 1986, which supplied a practical way to train networks with hidden layers and thereby answered the limits of single-layer perceptrons that Minsky and Papert had emphasized in 1969. The word connectionism has older roots in psychology, and Jerry Feldman and Dana Ballard used connectionist in 1982, but the PDP volumes made it a banner.
Distributed representation is the idea that gives connectionism its distinctive character: a concept is encoded as a pattern of activity across many units, and each unit participates in many concepts. There is no single cell for grandmother, and no symbol at a memory address. Geoffrey Hinton's 1986 chapter on the subject argued that such codes generalize naturally, because similar inputs produce overlapping patterns, they degrade gracefully when units are damaged, and they allow a network to store many items in the same weights. The cost is opacity, since one cannot read a rule off the weights. Rumelhart and McClelland's 1986 past-tense model illustrated the promise, learning English verb forms without explicit rules, and provoked a long dispute, including a detailed 1988 criticism by Steven Pinker and Alan Prince.
The sharpest challenge came in 1988 from Jerry Fodor and Zenon Pylyshyn in Cognition, in their paper Connectionism and Cognitive Architecture: A Critical Analysis. They argued that human thought is productive, systematic, and compositional: anyone who can think that John loves Mary can think that Mary loves John, and complex thoughts are built from reusable parts according to structure. A classical architecture explains this by operating on structured symbols, they said, while a network of association strengths has no principled reason to behave this way, so at best a connectionist model might implement a classical one. Replies denied that the premises were as strong as claimed, or that networks could not represent structure, and Paul Smolensky's 1988 paper proposed that networks operate at a subsymbolic level whose higher-level approximations resemble symbols. The dispute over whether networks can handle systematic, compositional generalization has never fully ended.
Connectionism's relation to deep learning is direct descent with a change of emphasis from cognitive modeling to engineering. The ingredients of the 1980s, hidden layers, backpropagation, and distributed representations, are the core of present networks, and learned word embeddings are distributed representations in the very sense Hinton described. What changed was scale: orders of magnitude more data, computation, and depth, together with architectural ideas such as convolution and attention. The shift also changed the goal, from explaining how people think to building systems that work, which is why many current practitioners use the tools without adopting connectionism's claims about the mind.
Two misconceptions recur. One holds that connectionism simply means artificial neural networks, but the term properly names a theoretical position about cognition, and a network used for ad-click prediction commits to nothing about minds. The other holds that connectionism is a literal model of the brain, though the PDP authors framed their units as neurally inspired abstractions, and backpropagation in particular has no established counterpart in cortical tissue. Equally, the Fodor and Pylyshyn objection is sometimes treated as refuted by the success of large networks, yet researchers continue to test language models on compositional and systematic generalization and to find uneven results, which keeps the old question alive in measurable form.
Read connectionism as the claim that intelligence lives in the weights of many simple interacting units: the 1986 program that survived its symbolic critics not by winning the argument outright but by turning into the engineering foundation of modern AI.