A new approach for neural network training and evaluation with applications to sensing systems
Xinjia Chen, Ernest Walker · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2011
Artificial neural networks are widely used in pattern recognition for sensing systems and other areas. In this paper, we propose to improve the performance of neural networks from the perspectives of output encoding rules, determination of training sample sizes, training performance index and evaluation of generalization error. We propose a new output encoding rule which significantly reduces the training error as compared to classical output encoding methods. Moreover, we develop a new training performance index which is closely relate to the generalization error and is a smooth function suitable for optimization by virtue of nonlinear programming. Furthermore, motivated by the crucial impact of training sample size on the generalization error and the computational complexity of training, we propose a rigorous method for determining appropriate number of training samples. Since the development of a neural network requires many cycles of training and performance evaluation, we introduce adaptive methods for estimating the generalization error. The new techniques of neural network training and evaluation have potential to improve the power of modern sensing systems.