Algorithm

An algorithm is a general solution procedure made of a limited sequence of clear steps. A sorting algorithm, for instance, describes how to bring unordered values into order; a concrete program is then one possible way of implementing that idea in a specific programming language.

Machine learning has its own kind of algorithm: a training algorithm — a fixed procedure applied to training examples that produces a model from them. Unlike classic algorithms, the result here isn’t an immediately readable answer, but a set of tuned parameters.

An example: The procedure of a simple spam filter: compare the sender with the address book, read the subject, count the exclamation marks, decide. You could write these steps on paper, and a person could work through them just as well as a computer. A cooking recipe is a similar picture, with one limit: it may say “a pinch of salt”, whereas an algorithm has to spell out every step unambiguously.

Not to be confused with a platform’s “algorithm”: When people say “the algorithm” decides what you see in a feed, they usually mean a whole recommendation system. It can combine several algorithms, fixed rules and trained models. Strictly speaking, an algorithm is only the procedure, not the model a training algorithm produces.

Where you’ll come across it: In math and computer science classes, for example with sorting or long division, in reporting on social networks and search engines, and wherever there’s talk of a procedure that selects, ranks or calculates something.

Explained in more depth in Program, Algorithm, Model Compared.