An early 1960s learning machine had a 'goof' button for mistakes
Raytheon's Cybertron, built by the early 1960s, stored its memory on punched tape and learned to recognise sonar signals, heart traces and speech through repeated coaching by a human teacher. When it got something wrong, the operator pressed a button marked 'goof', prompting it to reconsider. It was an early glimpse of machines that improve from feedback.
Machine learning is the branch of artificial intelligence concerned with algorithms that learn patterns from data and apply them to new cases without being explicitly programmed. The term was coined in 1959 by Arthur Samuel of IBM, whose 1950s checkers program estimated each side's chances of winning; 'self-teaching computers' was a rival name. Tom Mitchell later gave a crisp definition: a program learns if its performance on a task, by some measure, improves with experience. That practical framing echoes Alan Turing, who replaced the question of whether machines think with whether they can convincingly imitate a person.
The ideas drew on brain science. In 1949 the Canadian psychologist Donald Hebb proposed that connections between nerve cells strengthen through use, a principle mirrored by artificial neurons that adjust their links according to data. Walter Pitts and Warren McCulloch had earlier offered the first mathematical model of a neural network.
The path was not smooth. As AI embraced logic and knowledge-based expert systems, which dominated by 1980, statistical learning drifted out of the field into pattern recognition and information retrieval, and computer science abandoned neural networks. A handful of scientists based in other fields kept the approach alive, among them Geoffrey Hinton, David Rumelhart and John Hopfield, and their breakthrough arrived around 1985, when backpropagation was rediscovered.
Deep learning later swept the field. In 2012 Hinton's team, with Alex Krizhevsky and Ilya Sutskever, won the ImageNet competition by a wide margin with AlexNet. Generative adversarial networks followed in 2014, and in 2016 AlphaGo became the first program to beat a professional Go player without a handicap on a full board. The dominance since then is such that people often treat machine learning and AI as the same thing, though they are not.
Source: Machine learning
