PERSON

Ronald Fisher

The statistician who gave science its tools for honesty about uncertainty—maximum likelihood, analysis of variance, the design of experiments, the p-value—and who, by applying his own optimization framework to human beings, left the age of AI its most instructive and most uncomfortable inheritance.

Ronald Aylmer Fisher is the invisible founder of artificial intelligence: every learning system in production today runs, at its mathematical heart, on principles he worked out at an agricultural research station north of London before the Second World War. The training objective of a modern neural network is the maximum likelihood estimator Fisher named and justified in the 1920s. The evaluation of AI systems against benchmarks is the design of experiments he invented in 1935. The bias-variance trade-off at the center of machine learning is Fisher's variance decomposition, renamed. Even the p-value abuse driving the AI replication crisis is the abuse Fisher spent decades warning against. To understand how machines learn from data—and how they fool us—is to read Fisher, the man who built the rigorous discipline of inference under uncertainty that every learning system inherits. But Fisher is also the hardest subject this series has taken on. He was a lifelong eugenicist

Keep reading with YOU ON AI