Jay McClelland on Neural Networks and the Mind

McClelland describes his view that neural networks can connect the study of biology with questions about the mind. He recalls rejecting the separation of cognition from physiology and introduces Descartes’ mechanistic account of animal behavior.

The conversation considers Darwin’s struggle with the implications of evolution, evidence for continuity between humans and other animals, and the development of the brain. McClelland notes similarities between early visual pathways in humans and other animals.

The speakers discuss what may distinguish human cognition, the relationship between language and social intelligence, and whether a small evolutionary change could open the way to further development. They connect punctuated equilibrium with stages in child development.

They discuss the assembly and learning capacities of the brain, then turn to McClelland’s early career. He recalls joining UC San Diego, the influence of Explorations in Cognition, and discovering neural-network models of perception and memory.

McClelland recalls realizing that neural networks could help address questions about the mind, and describes the researchers who gathered around these ideas. The discussion introduces the meaning of parallel computation: many simple, autonomous units working at once.

The speakers explain how many interconnected units can process inputs simultaneously, using convolutional neural networks and visual recognition as examples. McClelland describes how layers contribute to a classification through distributed activity.

McClelland describes Rumelhart’s upbringing, mathematical psychology training, and early interest in building systems that could understand stories. He gives an example of the inferences needed to work out what a character intends.

The conversation moves from symbolic representations toward Rumelhart’s model of reading. McClelland explains how visual features, letters, words, and meaning can influence one another through a parallel, bidirectional process.

McClelland describes how he and Rumelhart replaced separate “experts” with neuron-like processing units and connection weights. The model allows context and visual input to jointly shape letter and word perception, illustrating the connectionist view.

The speakers compare neural-network recognition with human recognition and discuss why the computations behind a result may not be available to introspection. They ask whether connectionist systems can contain the depth of human knowledge.

McClelland explains why he prefers “radical emergentist connectionist” to “eliminative connectionist.” Using sand dunes as an analogy, he argues that higher-level thoughts can be real while arising from lower-level interactions.

McClelland reflects on Rumelhart’s death and introduces semantic dementia, a progressive loss of meaning across sensory and language experience. He describes how a patient’s ability to distinguish familiar categories can deteriorate.

McClelland explains how semantic knowledge can break down across pictures and words, and relates this to distributed representation. He describes Rumelhart’s decline and considers what the condition did—and did not—leave intact.

McClelland recounts spending time with Rumelhart as some abilities remained despite profound impairment. He then returns to the PDP research group and describes Hinton’s advice to derive learning rules from the task a network must solve.

The discussion explains objective functions, error, gradient descent, and Rumelhart’s extension of the delta rule to hidden layers. McClelland describes how error signals were propagated backward through a network, giving backpropagation its name.

McClelland recalls Hinton’s early work on neural networks and recursive computation, and describes his use of visual and geometrical explanations. The conversation focuses on how intuitive insight can accompany explicit understanding.

The speakers discuss how collaboration can shape scientific thinking, Hinton’s interest in reasoning, and the Boltzmann machine’s connection to logic and probability. McClelland introduces computational intelligence as a field that can study mechanisms beyond human biological limits.

McClelland describes mathematics as a way to explore idealized worlds and derive precise consequences. He uses triangles and their applications to explain how mathematical ideas can connect abstract objects with the physical world.

The conversation continues with mathematics’ practical uses, including counting and commerce. McClelland distinguishes notation from the ideas it expresses and explains how intuition can support new mathematical insights.

McClelland discusses the interplay between intuitive discovery and formal proof, and considers whether language models may produce novel outputs from training. The speakers also discuss AlphaZero, chess evaluation, and search.

The speakers consider examples of apparent machine insight and how formal systems developed historically. McClelland suggests that the capacity for abstract formal thought can be acquired through immersion, and asks how mathematical and linguistic cognition differ.

McClelland discusses how formal training shapes intuitions, drawing on Chomsky and research on linguists’ judgments about meaning. The conversation turns to beliefs, enculturation, and the limits of experts’ awareness of what needs to be explained.

The speakers examine how expertise can make ideas seem self-evident, using natural numbers as an example of knowledge once treated as innate but now understood as culturally constructed. McClelland considers how teaching and immersion affect learning.

The conversation considers whether expertise narrows what people can see and whether people can step outside their own habits of thought. McClelland describes research suggesting that explanations of choices may be post hoc and shaped by culture.

Asked for advice, McClelland describes his early uncertainty about a career in psychiatry and how his interests came together during his college years. He recounts the campus events and studies that helped him find a direction.

McClelland explains that enjoyment of statistical analysis and experiments drew him toward psychology. The speakers discuss noticing intrinsic interests, sustained immersion, and how personal inclinations and experience can work together.

The speakers discuss the role of timing and context in careers and their early interest in psychiatry. McClelland recounts his experience on a mental-health advisory council and criticizes an approach focused on medication as the route to curing schizophrenia.

The conversation compares psychiatry with computational and cognitive approaches to studying the mind. McClelland says he fears cognitive decline more than death, reflecting on the loss of his capacity to engage with ideas and collaborators.

McClelland discusses how awareness of limited time affected his decision to work on mathematical cognition. Asked about his legacy, he points to collaborative moments when ideas became clearer and contributed to scientific progress.

McClelland describes the satisfaction of developing ideas with collaborators and recounts being advised to remove theory from an experimental paper. He explains how he pursued theoretical work anyway and advises against letting reviewers or professional labels limit a person.

Asked about the meaning of life, McClelland says people make meaning for themselves, drawing on earlier traditions and their local circumstances. The speakers reflect on humans as emergent beings who create and pass on stories.

McClelland thanks the host for the conversation. The episode closes with a recorded statement from Jeffrey Hinton about curiosity-driven research and breakthroughs arising from work people are excited about.