# Connectionism Connectionism is the program in [[Cognitive_science|cognitive science]] and [[Artificial_intelligence|artificial intelligence]] that models mental processes as the [[Emergence|emergent]] behavior of large networks of simple, neuron-like units, where knowledge lives in the pattern of connection weights rather than in stored symbols and rules. Descended from [[Cybernetics|cybernetic]] models of the 1940s — [[Warren_Sturgis_McCulloch]] and [[Walter_Pitts]]'s 1943 logical neurons above all — and industrialized by today's [[Neural_network_(machine_learning)|neural networks]] in [[Machine_learning|machine learning]], it stakes a philosophical claim as much as an engineering one: that cognition is a [[Dynamical_system|dynamical]], statistical, brain-style process, and that the crisp symbol structures of classical [[Computer_science|computer science]] are at best a description of the network's aggregate behavior, not its mechanism. ## Two equations and a philosophy of mind Almost everything in classical connectionism follows from two lines. A unit's activation is a [[Nonlinear_system|nonlinear]] squash of its weighted input, aᵢ = f(Σⱼ wᵢⱼ·aⱼ); learning is small [[Feedback|error-driven]] weight change, Δwᵢⱼ = η·δᵢ·aⱼ, with learning rate η typically ≪ 1. From these come the school's signature properties: representation is distributed across many weights, so no single connection means anything; similar inputs yield similar outputs, giving free generalization; and damage degrades performance gracefully instead of catastrophically — all traits the [[Nervous_system|nervous system]] shows and rule tables lack. Cognition becomes trajectory-following in a high-dimensional state space, which puts the field nearer to [[Dynamical_systems_theory|dynamical systems theory]] and [[Statistical_mechanics|statistical mechanics]] than to logic, and makes [[Probability|probabilistic]] and [[Statistics|statistical]] reading of behavior native rather than bolted on. ## From logic neurons to the perceptron winter [[Warren_Sturgis_McCulloch|McCulloch]] and [[Walter_Pitts|Pitts]] proved in 1943 that nets of threshold units can compute any finite logical function — the result that led [[John_von_Neumann]] to describe computers in neural vocabulary. Donald Hebb's 1949 postulate gave learning its slogan (co-active cells strengthen their connection), and Frank Rosenblatt's perceptron (1958) added a convergence-guaranteed training [[Algorithm|algorithm]] for single-layer classifiers. Marvin Minsky and Seymour Papert's *Perceptrons* (1969) then proved exactly what one layer cannot do — parity and connectedness being the famous casualties — and funding followed the theorem: through the 1970s the symbolic program of [[Herbert_A._Simon]] and Allen Newell, running on rules and [[Expert_system|expert systems]], owned [[Artificial_intelligence|AI]], while network research survived at the margins in [[Computational_neuroscience|computational neuroscience]] and adaptive [[Signal_processing|signal processing]]. ## The 1986 revival Three results rebuilt the field. John Hopfield (1982) showed that symmetric recurrent nets minimize an energy function, so memories can be stored as [[Attractor|attractors]] — content-addressable recall as relaxation, with [[Multistability|multistability]] doing the work, and the mathematics borrowed from spin systems in [[Statistical_mechanics|statistical mechanics]]. The Boltzmann machine (Ackley, Hinton, Sejnowski, 1985) imported [[Ludwig_Boltzmann|Boltzmann]]-distribution sampling to learn internal representations stochastically. And the two-volume *Parallel Distributed Processing* (Rumelhart, McClelland, and the PDP group, 1986), with the backpropagation paper of Rumelhart, Hinton, and Williams the same year, made multi-layer training practical, dissolving the 1969 objection. The movement's provocation was the past-tense model: a single network that produced both regular and irregular English verb forms, plus child-like overregularization errors, without containing any rule — a direct shot at symbolic accounts of language, answered sharply by Pinker and Prince in 1988. ## The systematicity war Jerry Fodor and Zenon Pylyshyn's 1988 critique remains the canonical objection: thought is systematic and compositional — whoever can think that John loves Mary can think that Mary loves John — and classical symbol systems explain this for free, while networks must be specially trained into it. Connectionists answered along two lines: Paul Smolensky's subsymbolic position, on which symbols are coarse, approximate descriptions of fine-grained vector dynamics, and implementationalism, on which networks realize symbol systems the way [[Electrical_engineering|circuitry]] realizes [[Algorithm|algorithms]] — a truce in which levels of description coexist, as in the rest of [[Systems_science|systems science]]. The argument has never fully closed; it resurfaces every time a large model composes fluently while failing some crisp generalization a grammar would guarantee. ## The 100-step argument and neural plausibility Connectionism's empirical bet is anatomical. Jerome Feldman and Dana Ballard's 1982 timing argument notes that neurons operate on the order of milliseconds while recognition completes in a few hundred milliseconds, leaving time for only ~100 serial steps — far too few for long symbolic programs, but ample for massively parallel constraint satisfaction across the brain's roughly 10¹¹ neurons and 10¹⁴ synapses. The claim was never that units are faithful neurons: real [[Neural_circuit|neural circuits]] spike, obey Dale's law, and learn by mechanisms backpropagation does not respect, which is why [[Systems_neuroscience|systems neuroscience]] and [[Computational_neuroscience|computational neuroscience]] treat connectionist nets as [[Mathematical_model|models]] of computation at the level of populations, and why biologically plausible credit assignment is still an open research line rather than a solved footnote. ## Deep learning and the argument's afterlife The 2012 ImageNet result made deep networks the default technology of [[Machine_learning|machine learning]], and scale did what argument could not: [[Neural_network_(machine_learning)|neural networks]] now write, translate, and reason well enough that the old debate has inverted — the live question is no longer whether networks can exhibit structured cognition but whether symbolic description captures what they in fact do. The intellectual through-line back to [[Cybernetics|cybernetics]] is direct: [[Norbert_Wiener]]'s adaptive filters, [[Claude_Shannon|Shannon]]-style [[Information_theory|information]] accounting of what weights store, and [[Self-organization|self-organizing]] feature maps all read today as early chapters of one continuous field. Within this vault, connectionism is the bridge between [[Biocybernetics|biocybernetics]] and engineered [[Complex_system|complex systems]]: the standing demonstration that regulation, memory, and inference can be properties of a [[Graph_theory|graph]] and its weights — of organization as such — rather than of any part. **On the spine:** [[Neural_network_(machine_learning)]] · [[Cognitive_science]] · [[Cybernetics]] · [[Attractor]] · [[Machine_learning]] · [[Computational_neuroscience]]. ## Wikipedia : Wikitube **Strict pair:** [Wikipedia](https://en.wikipedia.org/wiki/Connectionism) : [Wikitube](https://en.wikitube.io/wiki/Connectionism) ## Previous hub tags Hubs: `Systems`. Portals: [[PORTAL_Cybernetics]], [[PORTAL_Decision_theory]], [[PORTAL_Information_theory]], [[PORTAL_Control_theory]], [[PORTAL_Emergence]]. --- *Repopulated 2026-08-12 · redlink fill · 0 deletions.*