Input Layer Optimization of Neural Networks by Sensitivity Analysis and its Application to Recognition of Numerals

Abstract

A major problem in applying a multi-layered neural network to image processing is how to select input components from among a number of features or pixels extracted from a recognized object. This is especially important for achieving both high speed processing and a high recognition rate.
This paper proposes a method of optimizing the number of input components, i. e. that of input neurons without lowering a recognition rate. In general, neural networks learn a recognition algorithm from input patterns and their desired output patterns. But if some input components are redundant, namely they are expressed by other input components or they do not contribute to recognition, their effect on the recognition algorithm, or their sensitivity to outputs is considered to be very small. Therefore, by analyzing input-output sensitivity of neural networks, redundant input components can be deleted. Repeatedly applying this method, the useful input components for recognition can be selected. For an experimental neural network for recognition of numerals with 12 feature input components, the input components could be reduced to eight without lowering a recognition rate.