Classification of handwritten digits and Japanese Kanji

Abstract

Dystal is an artificial neural network that associatively learns to classify handwritten zip code digits and Japanese Kanji characters. As a consequence of biologically motivated learning rules, Dystal has a number of mathematical properties such as a theoretical storage capacity of b/sup n/ nonorthogonal memories, where b is the number of discrete values and n is the number of output neurons, and a computational complexity of O(N). A brief overview of the network and results from character recognition experiments are presented. Dystal correctly classified 95% of previously unseen handwritten digits both with binary input (the original patterns) and with continuous-valued input (preprocessed versions of the original patterns). Dystal also was trained to classify handwritten Japanese Kanji characters and achieved a performance level in excess of 89% correctly classified. >