Was there machine learning before the deep-learning era?

Alongside the rule-writing tradition ran a different idea: let the program learn from experience rather than be told every rule in advance. Arthur Samuel, working at IBM, is the clearest early example. He “coined the term machine learning in 1959,” and is “most known within the AI community for his groundbreaking work in computer checkers in 1959.”^[1] His program got better by playing “thousands of games against itself,” which is recognizably the same idea behind modern learning systems.^[1]

The mechanism behind that improvement was concrete, not mystical. The program searched a tree of reachable board positions and picked its move with a minimax strategy, always assuming the opponent was also trying to win.^[1] To judge a position it used a hand-built scoring function weighing things like piece count, king count, and how close a piece was to being kinged, and with memory too limited to search every line to the end, Samuel used what is now called alpha-beta pruning to skip branches that could not change the outcome.^[1] The “learning” part was rote memory: the program remembered the value of every position it had already seen, and after each round of self-play it adjusted its own scoring function based on how the games actually turned out.^[1]

The other early learning thread was the perceptron, a simple model loosely inspired by a brain cell that adjusts itself from examples. This is where the two traditions collided. In 1969 Marvin Minsky and Seymour Papert published a book, Perceptrons: An Introduction to Computational Geometry, that proved sharp mathematical limits on what a single-layer perceptron could compute.^[2]

The book’s reputation outran its math. It “is often thought to have caused a decline in neural net research in the 1970s and early 1980s.”^[2] During that stretch, Wikipedia notes, “neural net researchers continued smaller projects outside the mainstream, while symbolic AI research saw explosive growth.”^[2] The learning approach did not die; it waited. Its revival is the story of the next topic, the deep-learning revolution.