Connectionism holds that intelligence is not programmed rule by rule but arises from learning in networks of simple units. The name reflects the idea that knowledge resides in the weighted connections between units (M. Mitchell, 2019, ch. 2): there is no place where the rule is written down, only a distributed pattern of weights.
Buckner (2024) distinguishes three waves (2024, pp. 119–121). The first runs from the McCulloch–Pitts neuron to Rosenblatt's Perceptron, and ebbs when Minsky & Papert (1969) show that a single layer of weights cannot learn functions such as XOR. The second arrives with backpropagation, popularized by Rumelhart et al. (1986), which makes it possible to train networks of three or four layers (Multilayer perceptron (backprop)); its manifesto was the two-volume Parallel Distributed Processing (1986) by Rumelhart, McClelland and colleagues (M. Mitchell, 2019, ch. 2). That wave brought ideas into cognitive science that were then revolutionary—distributed representation, soft constraints, gradual learning, graceful degradation under damage, human-like errors such as overgeneralization—and the third wave, deep learning, inherits them all (2024, p. 121).
Philosophically, connectionism is a bet about the architecture of the mind. Smolensky (1988) proposed a "subsymbolic" level between neurons and symbols; Fodor & Pylyshyn (1988) replied with a dilemma: either networks fail to explain the systematicity of thought, or they explain it because they implement a classical architecture, in which case they contribute nothing of their own as a theory of cognition (Millière & Buckner, 2024a, sec. 3.1). Andy Clark summed up the contrast with GOFAI in a slogan: subsymbolic systems are "bad at logic, good at Frisbee" (M. Mitchell, 2019, ch. 2).
Language models reopened the account. For Santelli (2026), the debate has become asymmetric: symbolists raise objections in principle, while connectionists build architectures that seem to do what should not be possible. But performance is not explanation. Knowing what a network has learned requires interpretability, and recent work suggests that its vectors implicitly realize symbolic structures (McCoy et al., 2026). Does that vindicate connectionism, or confirm the second horn of the dilemma?