An adaptive character recognition machine



Journal Title

Journal ISSN

Volume Title



High speed digital computers perform arithmetic and logical functions at extremely fast rates and line printers, plotters, and cathode ray tube displays have been developed which can output results at comparable rates. The link in the man-machine interface which is now receiving much attention is the computer input. The speed of this link could be greatly increased by a character recognition machine which could convert hand lettered programs and data directly into computer input eliminating manual card punching. One machine which appears to be applicable is a learning machine called the perceptron. This machine can be trained to distinguish between different optical patterns in a manner similar to the learning process of humans. The original perceptron built by F. Rosenblatt at Cornell Aeronautical Laboratory used a large array of photo-cells onto which the patterns to be recognized were projected. The outputs of these receptors were connected at random to the inputs of a set of fixed threshold gates whose outputs were connected through adjustable gain amplifiers to adaptive threshold gates. Training was achieved by adjusting the gain of the amplifiers and the thresholds according to a training algorithm. In order to investigate the feasibility of the perceptron as a practical hand lettered alpha-numeric character recognition machine a digital version has been designed and constructed using binary numbers stored in a core memory and an accumulator in place of the adjustable gain amplifiers and threshold gates of the original perceptron. Although this machine has severe limitations due to its size, tests run on the machine indicate that the concept is feasible and for any given capability the digital machine would be smaller, and more easily trained than the original analog version.