Neural Network Associator for Images and Their Chinese Characters

Abstract

A neural network with two layers of processing elements (PEs) is investigated to associate a group of 48x48 gray scale images with the corresponding 16x16 Chinese character bitmap images. The input layer has 2304 (48x48) PEs, and the output layer has 256 (16x16) PEs. In order to reduce the size of the weight matrix, each input PE only connects to two output PEs, for a total of 4608 connections. Each output PE connects to 18 input PEs (4608/256=18). The connections are assigned randomly, and remain unchanged throughout this the investigation. The momentum based back propagation learning algorithm is used to train the weight matrix. The system is capable of perfectly learning input-output associations when inputs are orthogonal to each other. However, the investigation finds that close to perfect learning still occur for a random set of images.