Pattern recognition and modern computers
Selfridge defined pattern recognition as the extraction of significant features from a background of irrelevant detail, and described how a computer could extract them: sequences of simple local image operations, after which the remaining blobs are counted.
Why it matters
Instead of a template for every shape, the machine filters the image in a few local steps and compares the resulting numbers with stored distributions for each symbol; the paper also sketches how the machine could search for good sequences itself.
The talk precedes both the perceptron and the same author's Pandemonium. Selfridge tied significance to context and experience, and said that the usual line between machines and men, that men can learn what is significant and machines cannot, was one he did not think valid. His model applied a few local image operations, an averaging one and an edge-emphasising one, in sequences, counted the blobs left and compared the counts with stored distributions for each symbol; he planned to have the computer hunt for good sequences itself. The framing returned in convolutional networks, where the features are no longer hand-specified.