The Master Algorithm
Pedro Domingos
The five schools of machine learning, for people who want the map before the maths.
Pedro Domingos wrote this before the current AI boom, and it holds up well because it's about ideas. His claim is that machine learning grew out of five separate schools, each with its own idea of what learning is. He thinks the big prize is one algorithm that combines all five.
He calls them tribes. The symbolists treat learning as logic, working backwards from examples to rules. The connectionists copy the brain, with networks of simple units whose connections strengthen with training, which is the ancestor of today's neural networks. The evolutionaries borrow from natural selection and breed better programs. The Bayesians treat everything as probability and update their beliefs as evidence comes in. The analogisers learn by spotting similarities, judging a new case by the old cases it most resembles.
He explains each one without equations, using stories and everyday examples, and he's upfront about each tribe's blind spots.
It's on the shelf because it gives the whole map. Almost everything people talk about now comes from one of those tribes, and it's easier to understand what a model can't do once you know which family it belongs to.
