Frank Rosenblatt

Frank Rosenblatt (1928-1971) was an American psychologist and computer scientist who invented the perceptron, the first trainable artificial neural network to be built and demonstrated, and who spent his short career arguing that machines could be taught to perceive by adjusting connections rather than by being given rules. Trained in psychology rather than engineering, he approached the problem as a theory of how brains might store information, and he died young, in a boating accident, leaving a body of work that was dismissed for a generation and then vindicated.

Rosenblatt earned his doctorate in psychology at Cornell University in the mid-1950s and then worked at the Cornell Aeronautical Laboratory in Buffalo, New York, where in 1957 he wrote the technical report The Perceptron: A Perceiving and Recognizing Automaton, supported by the US Office of Naval Research. The following year, his Psychological Review paper The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain (1958) presented the theory to a scientific audience. The model was built from simple units with weighted inputs and a threshold, and its connection weights were changed by a learning rule that corrected errors on training examples. He first simulated the idea on an IBM 704 computer; a public demonstration of it in July 1958, in which the machine learned to tell marked cards apart, drew national press coverage, including a New York Times story that predicted far more than the device could do.

The hardware version, the Mark I Perceptron, was built at Cornell Aeronautical Laboratory and demonstrated around 1960: a camera-like grid of 400 photocells fed a layer of association units, whose adjustable weights were physical potentiometers turned by electric motors during learning. It could learn to classify simple patterns such as letters and shapes presented to its photocell retina,. Rosenblatt wrote up the theory at book length in Principles of Neurodynamics: Perceptrons and the Theory of Brain Mechanisms (Spartan Books, 1962), which explored multi-layer variants, networks with feedback, and psychological applications, well beyond the single-layer device that history remembers. The perceptron convergence theorem, which guarantees that the learning rule finds a separating set of weights in a finite number of steps whenever the classes are linearly separable, was proved in several forms in this period, notably by Albert Novikoff in 1962.

The decline of the field after 1969 is usually pinned on Marvin Minsky and Seymour Papert's book Perceptrons, which proved that a single-layer perceptron cannot compute functions such as exclusive-or, and which, together with the era's hunger for symbolic artificial intelligence, shifted funding away from neural networks. Rosenblatt himself knew single-layer limits and had discussed multi-layer machines, but he had no efficient learning rule for the hidden layers, a gap that backpropagation filled in the 1980s. He died on 11 July 1971 in a boating accident on Chesapeake Bay, his forty-third birthday, while the perceptron was still out of favor; he did not live to see the revival.

The usual misconceptions about Rosenblatt are that his machine was a failed dead end, that Minsky and Papert refuted the whole idea of neural networks, and that the perceptron is a historical toy. In fact the perceptron's learning rule, its linear threshold unit and the language of weights and layers are direct ancestors of today's networks, and the IEEE Frank Rosenblatt Award, created in 2004, honors work in neural-network-inspired computing. What Rosenblatt did not have was data, computing power, and multi-layer training methods at modern scale, and his overreaching press-era claims in 1958 hurt his credibility. Still, the core claim, that a machine could acquire pattern recognition ability from examples by adapting its connection strengths, is the foundation of present-day deep learning.

Read Frank Rosenblatt as the psychologist-engineer who built the first learning neural machine decades too early: his perceptron was limited, but the idea of training weights from examples was exactly right.