In 1957–1958, Frank Rosenblatt at the Cornell Aeronautical Laboratory developed the Mark I Perceptron: a single-layer network that adjusted weights to classify patterns, implemented in hardware with photocells.
The press promised machines that would talk and walk. The perceptron learning algorithm, with guarantees for linearly separable classes (convergence theorem), was a theoretical and practical milestone.
Enthusiasm later collided with expressivity limits documented by Minsky and Papert. Even so, the Perceptron set the supervised-learning-from-examples pattern that dominates contemporary deep learning.
