In 1969 Marvin Minsky and Seymour Papert published “Perceptrons: An Introduction to Computational Geometry” with MIT Press. The book is a careful mathematical study of what single-layer perceptrons - the neural networks of the day - can and cannot compute. Its most famous result is that a single-layer perceptron cannot learn the exclusive-or (XOR) function, a fact the authors prove rather than assert. The book was not a hit piece; it was rigorous geometry. But its timing and authority landed hard on a young field that had been promising a great deal.
What followed is often called the first “AI winter” for neural networks. Research attention and funding drifted away from perceptron-style learning for years. The limitations the book identified applied to single-layer networks; multi-layer networks could in principle do more, but nobody yet had a practical way to train the hidden layers. The field largely waited.
The thaw is usually dated to 1986, when David Rumelhart, Geoffrey Hinton, and Ronald Williams published their account of backpropagation - a method for training the hidden layers of multi-layer networks by propagating errors backward through them. That work, together with the broader connectionist revival of the mid-1980s, reopened the path the 1969 book had seemed to close. Tellingly, the 1988 expanded edition of Perceptrons, still in print from MIT Press, adds a prologue and epilogue in which the authors discuss exactly this revival.
The story is not that Minsky and Papert were wrong - their proofs stand. It is that a precise statement of one architecture’s limits was read, for nearly two decades, as a verdict on a whole idea. The gap between “this specific model cannot do XOR” and “neural networks are a dead end” is where the seventeen years went.