The Master Algorithm

Domingos, a machine learning researcher, argues that the field consists of five distinct schools with different assumptions, techniques and philosophical origins, and that a unified algorithm combining all five is the objective the field should pursue.

The five tribes are the symbolists, who treat learning as inverse deduction and work with rules and logic; the connectionists, who reverse-engineer the brain and work with neural networks and backpropagation; the evolutionaries, who simulate natural selection through genetic algorithms; the Bayesians, who treat learning as probabilistic inference and update beliefs with evidence; and the analogisers, who reason from similarity and work with support vector machines and nearest neighbour methods.

For each he explains the core assumption, the master algorithm within that tradition, what it does well and where it fails. His argument is that each captures something real and that none is sufficient, and he sketches his own attempt at unification through Markov logic networks. He also covers what machine learning cannot do, the difference between correlation and causation in learned models, the data requirements, and the social consequences including the concentration of data and the effects on employment and privacy.

The book was published in 2015, before the transformer architecture and before the current scale of connectionist success. The subsequent decade has gone overwhelmingly to one tribe, which weakens the claim that unification is the path. Domingos has become a public critic of parts of the field since. The taxonomy remains the clearest available explanation of why different machine learning approaches exist and what each assumes.