History of Machine Learning

In 1952 Arthur Samuel of IBM built a checkers program that improved by playing against itself and adjusting its evaluation of board positions. By 1962 it beat a strong human player. Samuel coined the term 'machine learning' in 1959 to describe the approach: a computer getting better at a task without being explicitly reprogrammed.

The checkers program set the template that still defines the field — define a measurable objective, let the system adjust itself from data or self-play, and measure progress against human performance. Self-play in particular resurfaced decades later in systems like AlphaGo Zero.

In 1957 Frank Rosenblatt built the Mark I Perceptron, the first hardware implementation of an artificial neural network that could learn to classify simple visual patterns. Breathless press coverage promised machines that would walk, talk, and be conscious — and the backlash to that hype, crystallized by Minsky and Papert's 1969 critique, helped trigger the first AI winter.

The perceptron matters to machine learning history because it established learning from examples as a physical, working mechanism rather than a theory, and because its boom-and-bust reception set the pattern of overpromise and disillusionment the field would repeat.

Through the 1990s, machine learning quietly detached itself from symbolic AI. In 1992 Gerald Tesauro's TD-Gammon learned world-class backgammon through temporal-difference reinforcement learning and self-play, with no human strategy encoded. In 1995 Vladimir Vapnik and Corinna Cortes published support vector machines, a mathematically principled classifier that dominated practical machine learning for the next decade.

This was the era when 'AI' was a discredited label but machine learning thrived as applied statistics — spam filters, credit scoring, recommendation systems. The field learned to win on narrow, measurable problems while the grand ambitions waited for more data and compute.

In 2006 Netflix offered $1 million to any team that could improve its recommendation algorithm by 10%, turning machine learning progress into a public competition and popularizing ensemble methods. In 2009 Fei-Fei Li's team released ImageNet, over 14 million hand-labeled images, and its annual challenge became the benchmark that made computer vision progress measurable.

Together they demonstrated that large-scale data and shared benchmarks — not new algorithms alone — drive machine learning progress. ImageNet in particular became the launchpad for the deep learning breakthrough that followed.

In September 2012 AlexNet, a deep convolutional network trained on two consumer GPUs, won the ImageNet competition by an unprecedented margin. Traditional computer vision pipelines — hand-engineered features feeding a classical classifier — were abandoned almost overnight. In 2013 Word2Vec showed neural networks could learn meaningful vector representations of words at scale, bridging statistical NLP and the neural language models that led to transformers.

After 2012, 'machine learning' increasingly meant deep learning. The lineage from Samuel's self-improving checkers player runs directly through these systems to today's large language models, which are machine learning at the largest scale ever attempted.