“The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain” was published by Frank Rosenblatt in 1958 in Psychological Review (volume 65, issue 6, pages 386-408). Rosenblatt was a psychologist at the Cornell Aeronautical Laboratory, and the paper introduced the perceptron, the first artificial neural network that could learn.
Where the McCulloch-Pitts neuron of 1943 had fixed wiring, the perceptron could adjust itself. Rosenblatt described a unit that takes several inputs, weights each one, sums them, and fires if the total crosses a threshold. The key idea was the learning rule: when the perceptron makes a mistake on a training example, it nudges its weights in the direction that would have produced the right answer. Repeat this over many examples and the machine gradually learns to classify patterns it was never explicitly programmed to recognize.
This was new and genuinely exciting. Rosenblatt later built the Mark I Perceptron, a physical machine that learned to distinguish simple images, and the press reported it in grand terms - a machine that might one day walk, talk, and reproduce itself. The perceptron established the template for supervised learning from labeled data that still underlies most of machine learning.
The honest limit became famous. A single perceptron can only separate patterns that are linearly separable; it cannot learn even the simple XOR function. In 1969 Minsky and Papert published a book emphasizing these limitations, which is often blamed for cooling funding and enthusiasm for neural networks for years - the so-called perceptron freeze. The deeper fix, stacking perceptrons into multiple layers and training them with backpropagation, did not arrive in practical form until the 1980s.