Representation, Evaluation, Optimization

Definition

Domingos argues that despite their surface diversity, all learning algorithms — symbolist, connectionist, evolutionary, Bayesian, or analogical — reduce to the same three components: a representation (the formal language the learner uses to express hypotheses, which determines what it's even possible for it to learn), an evaluation function (a score, such as accuracy or likelihood, that distinguishes good hypotheses from bad ones), and an optimizer (a search procedure that finds the highest-scoring hypothesis the representation allows). "Hundreds of new learning algorithms are invented every year, but they're all based on the same few basic ideas."

In the Book

This decomposition is introduced early and used as an organizing lens for comparing the five tribes' algorithms across the book, then made explicit near the end of Chapter 9 and used directly in Chapter 10 for Domingos's argument that AI systems pose no autonomy risk: "the learner's representation circumscribes what it can learn... the optimizer then does everything in its power to maximize the evaluation function — no more and no less — and the evaluation function is determined by us." Because every learner, however powerful, is only as capable as this triple lets it be, and the evaluation function is human-set, Domingos concludes a learning system can't autonomously adopt goals outside what it was built to optimize — "a robot whose programmed goal is 'make a good dinner' ... can't decide to murder its owner any more than a car can decide to fly away."

Why It Matters

This is a general-purpose way to dissect any optimizing or search-based system, not just ML: ask what space of solutions it can even represent, what it's being scored against, and how it searches that space — and most disagreements about a system's behavior trace back to one of those three, not the others. It's also the concrete mechanism behind a recurring debate about autonomous AI risk: whether a system's goals can drift depends on whether its evaluation function is truly fixed by its designers or can itself become a target of optimization — a question this decomposition makes precise enough to argue about.