Recognition of Handwritten Digits in the Real World by a Neocognitron
Hayaru Shouno, Kunihiko Fukushima, M Okada · 2020
This chapter discusses a method of recognition using an artificial neural network called “neocognitron”. The neocognitron is a hierarchical neural network model of the primate visual system. The ability of the neocognitron to recognize patterns is influenced by the selectivity of feature-extracting cells in the networks. The chapter demonstrates that the neocognitron can robustly recognize a large set of patterns encountered in the real world. It examines the performance of the neocognitron with dual thresholds using a large database, ETL-1. The ETL-1 database is a handwritten character database that contains varieties of handwritten digits freely written by 1,400 people. The neocognitron consists of two types of cells: S-cells’ and ‘C-cells’. A layer consisting of S-cells is known as ‘S-cell layer’ and that of C-cells is known as ‘C-cell layer’. The chapter shows the relationship between thresholds and selectivity of S-cells. It also shows how the C-cells in the neocognitron respond when the thresholds of the S-cells are very high.