Marvin Minsky
Marvin Minsky (1927–2016) was an American cognitive scientist and mathematician, a co-organizer of the 1956 Dartmouth workshop and co-founder with John McCarthy of the MIT artificial intelligence effort, whose work ranged from building one of the first neural-network learning machines to co-authoring the critique of single-layer perceptrons and proposing the theory that the mind is a society of many simple agents. He won the ACM Turing Award in 1969.
He began with brains and hardware, not logic. Born in New York City in 1927, Minsky served in the US Navy, took a bachelor's degree in mathematics at Harvard in 1950, and earned his doctorate at Princeton in 1954 with a thesis on neural-analog reinforcement systems. In 1951, with fellow student Dean Edmonds, he built the SNARC (Stochastic Neural Analog Reinforcement Calculator), a machine of about forty vacuum-tube neuron units, with adjustable connections, that learned to run a simulated maze; it is commonly cited as one of the first artificial neural-network learning machines. His ideas on reinforcement in networks owed a debt to Hebb. He was also an inventor outside AI: in 1957 he patented the confocal scanning microscope, a design now fundamental to biological imaging.
The decade after Dartmouth made him a central organizer of the field. As a Harvard Junior Fellow he co-wrote the 1955 Dartmouth proposal, then moved to MIT in 1958 and, with McCarthy, started the MIT AI project in 1959, which grew into the Artificial Intelligence Laboratory (an independent laboratory from 1970, with Minsky as director until Patrick Winston succeeded him in 1972). His review "Steps Toward Artificial Intelligence" (Proceedings of the IRE, 1961) surveyed search, pattern recognition, learning, planning and induction as a unified research program. His students shaped early AI: Daniel Bobrow's STUDENT program for algebra word problems, Thomas Evans's geometric analogy solver, Terry Winograd's SHRDLU and Patrick Winston's work on learning structural descriptions, among others, each came out of his group's orbit. He also proved results on small universal Turing machines and wrote Computation: Finite and Infinite Machines (1967).
Perceptrons (MIT Press, 1969), written with Seymour Papert, is his most debated work. The book gave a mathematical analysis of what single-layer threshold networks can compute, showing that they cannot represent functions such as XOR or parity in general and that predicates like connectedness need units to see impractically large parts of an image. A popular legend holds that the book killed neural-network research on its own. The reality is more modest: it was a rigorous result about a restricted model, but the authors also voiced doubt that multilayer networks would be trainable, and combined with disappointing results and shifting funding it helped turn attention toward symbolic methods through the 1970s. The 1988 expanded edition added a defense of this reading. The later success of backpropagation did not overturn the theorems; it exploited the multilayer route the book had left open.
His later work was about representation and the architecture of mind. In 1974 he circulated A Framework for Representing Knowledge (MIT AI Memo 306), which introduced frames, structured packages of expectations with slots filled by defaults, a direct ancestor of object-oriented ideas and knowledge-representation languages, and a close relative of the scripts used in story understanding. The Society of Mind (1986), developed with Papert, argued that intelligence emerges from the interaction of many simple processes, agents, none of which understands anything by itself; its descendants include multi-agent and modular architectures. The Emotion Machine (2006) extended the picture to layers of reflective thinking. Minsky also spoke out against what he saw as the neglect of common sense, which he considered the main barrier to human-level machines.
The modern relevance is a mix of vindication and rebuke. Frames and semantic networks fed into the knowledge graphs and structured schemas that search engines and assistants use; the society-of-mind picture is echoed in mixtures of experts and in systems that orchestrate several specialized agents or models; SNARC anticipated learning from reward, now the basis of reinforcement learning. Yet deep learning succeeded largely through the single-algorithm, end-to-end, data-driven route he distrusted, and his insistence that common sense remains the hard problem still bites when systems fail on everyday reasoning. A pitfall is to describe him as simply anti-neural: his first machine was a neural network, and his criticisms were aimed at specific limits of a specific model. He died in Boston in January 2016.
Read Marvin Minsky as the field's restless theorist: a builder of one of the first learning machines, a rigorous critic of its first generation of successors, and a persistent advocate for the view that intelligence is an organization of many simple parts rather than one clever mechanism.