The ability of a neural network to generalize from a small set of handwritten digit training exemplars is dramatically improved with two new techniques. First, the number of neural network inputs is drastically reduced by using a log-polar coordinate system to produce a centered, constant size, constant average brightness image of 65 pixels which still retains sufficient information for discrimination. Second, generalized training examples are constructed from the training exemplars with carefully chosen random variations. The results of this work are impressive. The prior state of the art, Le Cun et al., used binary images, 784 inputs, 4635 nodes, 98442 connections, 9840 training exemplars, and required three days to train on a Sun SPARCstation 1. This work used 65 inputs, 75 nodes, 660 connections, 160 training exemplars, and required one hour to train on an AT-class PC, yet its results appear to be similar to those reported by Le Cun et al.
展开▼