McClelland, Jay

cognitive scientist, Stanford University; co-author with David Rumelhart of the Parallel Distributed Processing books

Jay McClelland: Neural Networks and the Emergence of Cognition

Lex Fridman Podcast · Sep 20, 2021

• Jay McClelland
• parallel distributed processing
• connectionism
• David Rumelhart
• Geoffrey Hinton
• backpropagation history
• interactive activation model
• semantic dementia
• radical emergentist connectionism
• mathematical cognition
• intuition and formal reasoning
• Mozart effect and cultural immersion

Core ideas

McClelland frames his life's work as an attempt to bridge biology and mind, rejecting the Cartesian-descended assumption (still present in 1960s cognitive psychology) that the nervous system is only "peripherally" relevant to understanding thought. He traces his own path to neural networks: as a young UCSD professor in the 1970s he was captivated by James Anderson's linear-algebra-based neural models of memory and categorization, and by a 1979-80 UCSD conference ("Parallel Models of Associative Memory") that brought together the community — including Geoffrey Hinton, whose PhD dissertation had impressed Rumelhart and Don Norman — that would eventually coalesce into the "PDP Research Group." He explains "parallel distributed processing" literally: each neuron-like unit is a simple, autonomous computational element operating simultaneously with millions of others, in contrast to the sequential, single-processor model underlying classical computer science and symbolic ("good old-fashioned") AI.

He gives an extended account of David Rumelhart as a collaborator and person — a mathematically gifted South Dakotan trained in mathematical psychology at Stanford who grew dissatisfied with symbolic AI's inability to explain how humans draw rich inferences (illustrated by his example of "Margie" hearing the ice cream truck), and who developed the "interactive activation model of reading," in which every level of linguistic interpretation (pixels, letters, words, meaning) simultaneously and bidirectionally constrains every other level. McClelland describes co-building with Rumelhart the interactive activation model of letter perception, replacing the model's original hand-coded "experts" with neuron-like units connected purely by weighted connections — a result he says convinced them both that connectionism could genuinely address cognitive questions. He calls himself (revising an earlier self-description as a "radical eliminative connectionist") a "radical emergentist connectionist": higher-level cognitive constructs are real and emergent, not illusions to be eliminated, but they don't exist as fixed, separable entities — comparing them to Douglas Hofstadter's metaphor of sand dunes, real and describable but fundamentally fluid products of lower-level dynamics.

He recounts how backpropagation emerged: Hinton, visiting UCSD on sabbatical in 1982 (alongside other collaborators including Francis Crick and Paul Smolensky), argued researchers should stop looking to biology for learning rules and instead define an objective function and use gradient descent to adjust connection weights to reduce error — building on existing single-layer gradient-descent methods (the Widrow-Hoff "delta rule"). Rumelhart generalized this to multi-layer networks via what he first called the "generalized delta rule," which required back-propagating error signals from the output layer to hidden layers — hence "backpropagation." McClelland describes Hinton as an unusually intuitive, geometrically-minded thinker (favoring hand gestures and drawings like a "ravine" metaphor for gradient descent over equations) whose early, largely unpublished 1970s-80s ideas anticipated concepts resembling transformers and recursive computation via fast weight changes, and situates Hinton's Boltzmann Machine work as an attempt to connect his family's Boolean logical lineage with probabilistic, graded constraint satisfaction.

On mathematical cognition, McClelland describes mathematics not as symbol manipulation but as a set of tools for reasoning with precision about idealized objects (citing Tristan Needham's analogy that studying math as pure symbolism is like studying music without ever hearing a note) and argues mathematical insight arises from an interplay between connectionist-style intuition (a sense that something is "obviously true") and formal, logical proof — quoting Poincaré's "it is by logic that we prove, but by intuition that we discover." He connects this to a broader hypothesis that formally trained thinkers (linguists, mathematicians, philosophers) develop a distinctive, culturally-acquired mode of "systematic" thought through sustained immersion — citing Lila Gleitman's finding that professionally trained linguists' semantic intuitions differ markedly from ordinary speakers' — and argues this immersion process, not necessarily special innate biology, plausibly explains cases like Chomsky's (raised in a household steeped in rabbinical/academic reasoning) or Mozart's (immersed in music from infancy), a pattern he dubs "the Mozart effect": the synergy of experience-driven immersion with whatever underlying resonance an individual brings to it. He connects this to the psychological finding (Nisbett and Wilson) that human self-reported explanations of our own behavior are frequently post-hoc rationalizations rather than accurate reports of the actual causal process — meaning even introspection about our own reasoning is itself culturally shaped rather than a transparent window into the mind.

McClelland reflects candidly and movingly on Rumelhart's decline from a progressive, still not fully understood neurological illness his own research group later helped characterize and name "semantic dementia" — a condition in which patients gradually lose the ability to differentiate related concepts (a monkey study patient eventually called all large four-legged animals "horses," small ones "cats," and mid-sized ones "dogs") in a way McClelland connects directly to how distributed representations degrade in PDP models. He describes visiting Rumelhart during his decline (recalling how he could still navigate to a restaurant and point to what he wanted on a menu despite being unable to name it), framing this as evidence that cognition is not monolithic but graded and multipartite, with partial competencies persisting independently of others — a scientifically fascinating but personally painful window into the very theories the two of them had built together. He closes discussing his greater fear of cognitive degeneration than of death itself, his sense of urgency (adopting harder, more important research questions roughly once a decade as he calculates his remaining productive years), and a view of meaning as something humans construct locally and personally rather than discover, emerging from particular, contingent collaborative contexts — his own exemplified by the "peculiar" convergence of ideas, people, and timing that produced the PDP research program itself.

Excerpts