Connectionism returns

This approach was called connectionism. Instead of writing rules for a machine to follow, you built a network of connected units that could adjust their connections based on what worked and what didn't. Show it enough examples, let it correct itself, and over time it learned.

It had failed before, or at least been declared a dead end. In 1969, Minsky and a colleague published a book showing real limitations of early neural networks, and most of the field took it as a verdict. But the researchers who stayed found ways around those limits.

In 1986, the researcher and colleagues published a technique for training deeper, more capable networks in a way that actually worked. They called it backpropagation: a method for telling each part of the network how much it contributed to a wrong answer, so it could adjust. The payoff was still decades away. But the direction had changed.