Its Own Operators

The typical pattern-recognition program is either elaborately preprogrammed to process specific arrays of input patterns, or else it has been designed as a tabula rasa, with certain abilities to adjust its values, or "learn." The first type often cannot identify large classes of patterns that appear only trivially different to the human eye, but that would completely escape the machine's logic (Bailey and Norrie, 1957; Greanias et al., 1957). The best examples of this type are probably capable of being extended to process new classes of patterns (Grimsdale et al., 1959a; Sherman, 1959). But each such extension would seem to be an ad hoc complication where it should be a simplification, and to represent an additional burden of time and energy on both programmer and computer. The latter type of self-adjusting program does not, at least as yet, appear to possess methods for accumulating experience that are sufficiently powerful to succeed in interesting cases. The random machines show relatively poor identification ability (Rosenblatt, 1958, 1960a). (One exception to this statement appears to be Roberts' modification of Rosenblatt's Perceptron (Roberts, 1960). But this modification appears to make the Perceptron an essentially nonrandom computer.) The most successful of this type of computer, to date, simply accumulates information or probabilities about discrete cells in the input matrix (Baran and Estrin, 1960; Highleyman and Kamentsky, 1960). But this is an unusually weak type of learning (if it should be characterized by that vague epithet at all), and