Using Prior Information to Improve the Approximation Performances of Neural Networks
Enachescu Calin, Theodore E. Simos, George Psihoyios, Ch. Tsitouras · AIP conference proceedings · 2007
Computers solve problems by using algorithmic approaches consisting of a set of rules that guides them to eventual solution of problems. This approach is different from neural computation that is based on neural networks and solves problems by learning and absorbing experience through the modification of their internal structure to accomplish a problem solution. The neural network's learning implies that the available information is usually divided into two categories: examples of function values otherwise known as training data and their corresponding prior information. This paper presents our approach to improve the learning performances of neural networks by using some additional information from the training data set through regression analysis. We discuss the theoretical foundation leading to the derivation of the prior information and also consider some experiments conducted to confirm our theoretical results based on the computation of Learning Error and Generalization Error metrics.