The XOR ProblemThe XOR Problem

XOR, the exclusive-or function, outputs true when exactly one of its two inputs is true. It is one of the simplest functions in logic. In 1969, Marvin Minsky and Seymour Papert made it the centerpiece of their book Perceptrons, a rigorous mathematical analysis of what Frank Rosenblatt's celebrated learning machines could and could not do.

Their result was clean and devastating. A single-layer perceptron can only learn functions whose positive and negative examples can be separated by a straight line, and XOR's four cases cannot be. No amount of training data or clever tuning could ever fix this, because the limitation was geometric, not practical.

Minsky and Papert acknowledged that networks with intermediate layers could represent XOR, but they were pessimistic that any procedure could reliably train such networks, and no good method was known at the time. The book landed amid rising skepticism about AI's grand promises, and funding for neural network research collapsed through the 1970s.

The episode became known as a key trigger of the first AI winter, at least for the connectionist camp. Rosenblatt died in 1971, and for years the field's center of gravity shifted decisively toward symbolic AI.

The irony is that the answer was hiding in the objection. Multi-layer networks trained by backpropagation, popularized by Rumelhart, Hinton, and Williams in 1986, learn XOR trivially, and the hidden layers Minsky and Papert doubted became the 'deep' in deep learning. The XOR problem endures as a lesson in how a correct proof about a narrow architecture can be misread as a verdict on an entire idea.