A new approach to hand-written character recognition

Abstract

A novel, biologically motivated, computationally efficient approach to the classification of hand-written characters is described. Dystal (DYnamically STable Associative Learning) is an artificial neural network based on features of learning and memory identified in neurobiological research on Hermissenda crassicornis and rabbit hippocampus. After a single pass through the training set, Dystal correctly classifies 98% of previously unseen hand-written digits. Similar training on hand-printed Kanji characters results in learning to read 40 people's handprinting of 160 characters to 99.8% accuracy (a task analogous to learning the latin characters in 40 different fonts) and reading different people's handprinting with 90% accuracy.