Textbook of AI

The Mark I Perceptron was the hardware implementation of Frank Rosenblatt's perceptron algorithm, built at the Cornell Aeronautical Laboratory in 1957–58 with funding from the US Office of Naval Research. Unveiled at a press conference in July 1958, it was the first physical machine to learn in the modern sense, adjusting its own parameters from examples to perform a pattern-recognition task , and is the direct ancestor of every neural-network accelerator that has followed.

Hardware

The Mark I had three layers:

1. A 20×20 grid of cadmium-sulphide photocells acting as a "retina", with 400 light-sensitive inputs.
2. 512 association units ($A$-units), each randomly wired to a small subset of the retina cells with fixed (random) weights, computing a thresholded sum.
3. Eight output units ($R$-units), fully connected to the association layer.

The adjustable weights between the association units and the response units were implemented as motor-driven potentiometers, small electric motors physically rotated a shaft to change a resistance, encoding learning as torque applied to a wiper. Updates followed the perceptron learning rule: when a training example was misclassified, weights on active inputs were nudged in the direction that would correct the output. The use of fixed random projections to the association layer, and trainable weights only at the final stage, anticipates a number of much later ideas, including extreme learning machines and reservoir computing.

Reception and overselling

The press demonstration was characteristically over-sold. The New York Times (8 July 1958) reported that "the Navy revealed the embryo of an electronic computer today that it expects will be able to walk, talk, see, write, reproduce itself and be conscious of its existence." The actual Mark I could be trained to discriminate between simple shapes (squares vs circles) and printed letters from various positions on the retina, useful but a long way from the Times's description. The mismatch between rhetoric and capability is an early instance of the AI hype cycle that has recurred throughout the field's history.

Significance

The Mark I demonstrated that:

• A learning algorithm could be embodied in physical hardware.
• Pattern recognition could be performed without hand-coded features (the random $A$-layer made any features it had).
• Convergence guarantees existed: Rosenblatt's perceptron convergence theorem proved that, for linearly separable training data, the algorithm would find a separating hyperplane in a finite number of updates.

Minsky and Papert's 1969 book Perceptrons showed the limits of the single-layer architecture, most famously its inability to learn XOR , and contributed to the funding collapse of the first connectionist wave. But the Mark I's lineage is unbroken: Rumelhart, Hinton and Williams' backpropagation paper (1986), LeNet (1989), AlexNet (2012), and the GPUs powering today's foundation models all sit downstream. The original Mark I now resides in the Smithsonian National Museum of American History.