Pandemonium
Selfridge described recognition as several layers of simple agents: some see features of the image, others assemble guesses from them, and a last one picks the loudest.
Why it matters
A layered feature hierarchy with learned weights is described here before there was any way to train one.
Weights in the scheme were to be tuned by outcome and unsuccessful detectors replaced. Selfridge gave no mechanism for training the deeper layers, and that missing part is exactly what back-propagation later supplied. The scheme influenced the neocognitron and later convolutional networks directly.