Bayesian networks

The AI of the 1970s had mostly avoided probability. It was thought too expensive to compute and too far from how people reason, and the expert systems had improvised with certainty factors and fuzzy logic instead. Judea Pearl's 1988 book ended the argument. A Bayesian network draws the variables of a problem as nodes and the direct dependencies between them as arrows; the graph makes the joint distribution tractable, and Pearl's belief-propagation algorithm passes messages along the arrows to update every belief when evidence arrives.

The book gave AI a rigorous way to reason under uncertainty and made the 1990s, in the mainstream conferences, the decade of probabilistic methods. Medical diagnosis, fault detection, spam filtering and the early Microsoft Office assistant ran on Bayesian networks. Hidden Markov models and Kalman filters turned out to be special cases.

Pearl went further in the 2000s into causality, asking what a graph must look like to support "what if" reasoning rather than just prediction, and won the Turing Award in 2011. His later complaint about deep learning, that it is curve-fitting without a model of cause, is the current form of the argument his book started.