The story
In 1958 Frank Rosenblatt, a psychologist at the Cornell Aeronautical Laboratory in Buffalo, New York, described the perceptron, a simple artificial neural network that could learn from examples. In July 1958 he showed a version running on an IBM 704 computer, and the Navy, which funded the work, announced it to the press. The New York Times wrote that it was the embryo of a computer that the Navy expected would be able to walk, talk, see, write, reproduce itself and be conscious of its existence, a claim that was far beyond what the machine could do.
The perceptron drew on two earlier ideas. In 1943 Warren McCulloch and Walter Pitts had described a mathematical model of a neuron that adds up inputs and fires if the total passes a threshold. In 1949 Donald Hebb had proposed that connections between brain cells strengthen when they are used together. Rosenblatt put these together. A perceptron takes inputs, multiplies each by a weight, sums them and outputs a yes or no. When it is wrong, the learning rule nudges the weights toward the right answer, and for problems that can be separated by a straight line the procedure is guaranteed to find a solution. Rosenblatt then built hardware, the Mark I Perceptron, with a grid of 400 photocells acting as an eye and motors that adjusted the weights, which learned to recognize simple shapes and letters.
The reaction swung between hype and disappointment. In 1969 Marvin Minsky and Seymour Papert published the book Perceptrons, proving that single-layer networks cannot solve some simple problems, such as the exclusive-or function. This contributed to a collapse in funding for neural network research in the 1970s. The approach came back in the 1980s with backpropagation, a method for training networks with several layers, and in the 2010s with deep learning on powerful hardware. Today's large language models and image systems descend from the perceptron.
Why it mattered
- The perceptron introduced the learning-by-adjusting-weights approach at the heart of all modern neural networks.
- Its limits, set out by Minsky and Papert in 1969, led to a long period of reduced funding for neural network research.
- Backpropagation in the 1980s solved the multi-layer problem and revived the field.
- Deep learning in the 2010s, which now powers speech, vision and language systems, grew from the same idea scaled up.
Sources
- Professor's perceptron paved the way for AI 60 years too soon Cornell Chronicle
- Early days of machine learning Smithsonian Magazine
Drafted by AI (Claude) from the sources above and checked against them by AI; no historian has reviewed this page. Spot a mistake? Suggest a correction. How pages are made
Community notes
Loading notes…