Parallel Distributed Processing
Rumelhart and McClelland's two volumes gathered connectionism into a coherent programme: representation is distributed, and knowledge lives in the weights rather than in symbols.
Why it matters
It was the book, not the paper, that made networks widespread: it gave a whole generation the vocabulary, the problems and the code.
The first volume sets out the theory, the second psychological and biological applications. It carries both the back-propagation chapter and a chapter on the Boltzmann machine. It sold out faster than the publisher expected. The word connectionism entered general use from this book as the name of a direction rather than of one model.