Machine learning

In conventional software, a programmer writes each rule, such as “flag any card payment over $5,000 made abroad”. In machine learning, the programmer chooses a model, which is a formula with adjustable numbers called parameters, and a training procedure that sets those numbers from data, such as past payments labeled as fraud or not fraud. Tom Mitchell’s textbook definition makes learning something that can be checked: “A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E.” The performance measure names what is counted. In Mitchell’s checkers example, the task is playing checkers, the measure is the percent of games won against opponents, and the experience is practice games the program plays against itself. For the fraud example, the measure could be the share of payments later confirmed as fraud that the model flags, counted on payments it did not see during training and reported together with the share of legitimate payments it flags by mistake.

Machine learning is a method rather than a kind of AI product: predictive AI, computer vision and generative AI are defined by what a system does, and each is usually built with machine learning. An MIT Sloan explainer (Sara Brown, 21 April 2021) says that “When companies today deploy artificial intelligence programs, they are most likely using machine learning”. An AI agent usually runs on a large language model, which is also trained from data. AI is the wider field: the EU AI Act lists machine learning approaches “that learn from data how to achieve certain objectives” alongside “logic- and knowledge-based approaches that infer from encoded knowledge” (Recital 12).

Arthur L. Samuel used the term in the title of Some Studies in Machine Learning Using the Game of Checkers (IBM Journal of Research and Development, July 1959). His program started from the rules of checkers and a list of parameters “whose correct signs and relative weights are unknown and unspecified”, and adjusted the terms and their weights as it played. Samuel wrote that “Programming computers to learn from experience should eventually eliminate the need for much of this detailed programming effort.” The paper notes that his first program with learning was completed in 1955 and demonstrated on television on 24 February 1956. Tom M. Mitchell gave the definition above in his textbook Machine Learning (McGraw Hill, 1997). A definition often credited to Samuel, “the field of study that gives computers the ability to learn without explicitly being programmed”, as in the MIT Sloan explainer, does not appear in the 1959 paper.