In classical programming someone writes the rules and the computer applies them. Machine learning reverses the order: the system is given data and expected answers, and it finds rules that connect the one to the other. Chollet (2021) puts it by saying that such a system is trained rather than explicitly programmed, and that three things are needed: input data, examples of the expected output, and a way of measuring how far the current output is from the desired one (2021, secs. 1.1.2–1.1.3). Learning, in this technical sense, is a search through a predefined space of possible transformations, guided by that error measure.
The name comes from Arthur Samuel, who in the 1950s programmed an IBM 701 to play checkers and improve with experience (M. Mitchell, 2019, ch. 9). Today the field ranges from a decision tree to GPT-3, and it is organized by the kind of signal available: supervised, unsupervised, self-supervised, or reinforcement learning. Its historical foil is symbolic AI (GOFAI), which relied on explicit rules written by programmers and experts (Chollet, 2021, sec. 1.1.1).
Does the machine really learn? Witten et al. (2017) propose an operational definition: things learn when they change their behavior in a way that makes them perform better in the future. But they immediately note that a comfortable slipper also meets it: it molds itself to the foot and performs better as a slipper, yet no one would say it learned (2017, p. 8). Their way out is pragmatic: the field goes by this name without presupposing any philosophical stance on what learning is (2017, p. 9). For a philosophical reader, that declared neutrality is exactly where the discussion begins (see training and generalization).