A recogniser that makes its own operators
In May 1961 Leonard Uhr and Charles Vossler described a pattern-recognition program for the IBM 709 that starts without any operator for processing its input and generates, evaluates and replaces them itself. The input is a character as a 20 × 20 matrix of zeros and ones, and the operators are 5 × 5 windows. With no pre-programmed operators it recognised 86 % of the characters of sets 2-7 of the five letters A-E, and 94 % on a second pass.
Why it matters
This is an early program in which not only the weights of features but the set of features itself is learned: weak operators are dropped and new ones generated in their place. The tests are small — five letters, 35 characters a run — and were made by the authors of the program. This is an editorial assessment.
The design. The program was written for the IBM 709, about 2,000 machine instructions. An unknown pattern is a 20 × 20 matrix of zeros and ones, framed by a mask. An operator is a 5 × 5 matrix of zeros, ones and blanks; it is slid over the pattern, and the matches give four characteristics on a 0-7 scale. Operators are formed by random methods and by combination; the characteristics are compared with the lists in memory, one for each known pattern, and the name of the nearest list is the answer. By success or failure the «amplifiers» of characteristics are turned up or down; operators with low amplification are dropped and new ones made. Time: about 25 seconds per character with 5 patterns and 40 seconds with 10. Tests (by the authors). 48 pre-programmed operators: 70-80 % on the last three or four of seven sets of five hand-printed letters A-E. Cells of the matrix and 500 combinatorial operators (900 characteristics): a little more than 30 %. All operators generated by the program itself (at most 40 matrix and 160 combinatorial): 86 % in sets 2-7 (the first set cannot be identified by construction), 94 % on a second pass over the same sets, with errors on two of 35 characters. Three passes over three sets of the first ten letters: 29 of 30 (97 %) on the third encounter and 70 % on a fourth, new set. Six line drawings (chair, table, two faces, two kinds of particle decay): after the first set 50 % in the second, after a second pass all of them. What the record does not claim: results on the whole alphabet (the authors say they did not run it because of machine time); the «about 80 % after two passes» for a modified faster program (the authors themselves call it preliminary); a comparison with the Perceptron on the same data; that the paper was reprinted in Computers and Thought (that comes from reference listings, not from what was read). The authors' own account without a second source: medium confidence.