Machine Learning
Machine learning is the study of computer programs whose performance on some task improves with experience, as opposed to programs whose behavior is fixed entirely by rules a programmer wrote out in advance. The term, which now names an enormous branch of computer science and industry, entered the literature through Arthur Samuel's 1959 checkers paper, and the two most famous ways of defining it, a loose one attributed to Samuel and a precise one from Tom Mitchell, are worth keeping carefully apart.
Arthur Samuel is conventionally credited with the term: his paper Some Studies in Machine Learning Using the Game of Checkers (IBM Journal of Research and Development, 1959) put it in the title and used it to describe programs that learn from their own play. The credit is customary rather than airtight, since nobody can certify that no earlier writer ever strung the two words together, and related notions circulated already: Alan Turing's 1950 paper Computing Machinery and Intelligence had proposed teaching a child machine instead of programming an adult mind, and Frank Rosenblatt's perceptron work of 1957 and 1958 described a machine that learned to classify patterns. What Samuel did was attach the phrase to a concrete, demonstrable program, in a venue read by the engineering community, at the moment when learning machines were drawing attention from both optimism and skepticism.
The famous slogan, that machine learning is the field of study that gives computers the ability to learn without being explicitly programmed, is widely attributed to Samuel but should be treated as a paraphrase of his meaning, not as a verified verbatim quotation from the 1959 paper. The paper does state his motivation: programming a computer to learn from experience should, he argued, eventually remove the need for much of the detailed programming effort. The quoted one-liner circulates in textbooks, lectures and corporate pages, and its exact original source is hard to pin down, so careful writers say attributed to Samuel or in the spirit of Samuel. It is also slightly misleading in itself, because a learning program is still explicitly programmed: the learning algorithm, the model class, the features or architecture and the training objective are all written by people. What is not hand-written is the particular behavior, such as which weights end up in an evaluation function.
A more rigorous formulation comes from Tom Mitchell's 1997 textbook Machine Learning (McGraw-Hill), which says a program learns from experience E with respect to a class of tasks T and a performance measure P if its performance at tasks in T, as measured by P, improves with experience E. The definition is useful because it forces three questions before anything can be called learning: what task is being attempted, how success is measured, and what experience is supplied. Mitchell illustrated it with checkers, tying the story back to Samuel: the task T is playing checkers, the performance measure P is the percentage of games won against opponents, and the experience E is playing practice games against itself. Applied to other cases, T may be classifying emails, P may be error rate on held-out messages, and E may be a labeled collection of past mail.
Machine learning is often confused with artificial intelligence as a whole, and with how people learn, though it is a subfield of the former and an engineering method rather than a model of the latter. Rule-based expert systems, logic programming and classical search are artificial intelligence without being machine learning; conversely, ordinary statistical regression, fitted to data, is machine learning in Mitchell's sense even though nobody calls it intelligent. Another pitfall is the idea that learning is automatically good: improvement on the measured P may not transfer outside the training distribution, and a model can learn shortcuts or biases hidden in its experience. Since the 1990s the field has leaned increasingly on probability and statistics, and in the 2010s on deep neural networks trained with large data sets, but Mitchell's three-part test still applies to each of them.
Read machine learning as a promise with a measuring stick: Samuel's phrase names the ambition of programs that improve from experience, and Mitchell's task, performance and experience triple says precisely what has to be shown before the word learning is earned.