Perceptrons (Minsky-Papert critique)

Editorial record

Plain-language summary

Through geometric and algebraic analysis the authors characterized exactly which predicates a single-layer perceptron can and cannot represent, showing that some require information the local, limited-order perceptron cannot combine, with XOR and connectedness as prominent examples of failures. They noted that multilayer networks could in principle overcome these limits but that no effective training method for them was then known. The rigorous negative results are widely credited with cooling early enthusiasm and funding for neural network research until multilayer training via backpropagation was popularized in the 1980s.