The Split: Neural Networks Fall from Favor after Perceptrons

Easily one of the most important and (in)famous moments in AI history, the moment when the field essentially was split in two. This division would last for decades and greatly impact the opposing side of learning almost irreparably. Into the 2000s, researchers in the symbolic camp would still cite (very mistakenly) Marvin Minsky for why neural nets would never work. Well this is the reason. By 2010, the learning side had all but won. By the second quarter of this century, deep learning, a direct descendant of Rosenblatt's perceptron, would be king.

Marvin Minsky and Seymour Papert’s Perceptrons (1969) entrenched the two artificial intelligence traditions, by seemingly disproving the learning side and Rosenblatt's perceptron.

Minsky and Papert backed symbolic AI's programs based on logical rules. Rosenblatt's learning tradition built networks that adjusted their connections from examples, whose limits the book set out in mathematical proofs. Those proofs were read as showing that neural networks had no future, and the Nobel Committee for Physics writes that the book “led to a hiatus funding-wise” for neural network research.

Sociologist Mikel Olazaran traces that reading to symbolic AI’s rise and its control of resources and authority. When neural networks revived in the late 1980s, the official history was rewritten. Yes, it would take two decades before anyone, especially in the West, took neural nets seriously again.