Classification of Numbers Using Neural Networks

Abstract

Pattern recognition in images is a classic application of neural nets. This chapter builds upon the previous one by exploring multi-layer networks, which fall into the Machine Learning branch of our Autonomous Learning taxonomy. In this case, we will look at images of computer-generated digits, and the problem of identifying the digits correctly. These images will represent numbers from scanned documents. Attempting to capture the variation in digits with algorithmic rules, considering fonts and other factors, quickly becomes impossibly complex, but with a large number of examples, a neural net can readily perform the task. We allow the weights in the net to perform the job of inferring rules about how each digit may be shaped, rather than codifying them explicitly.