FINITE-SAMPLE SIZE AND NEURAL MODEL UNCERTAINTY

Training a feedforward neural network leads to a model affected by unc ertainty. A measure of this uncertainty gives an important indication of neural model's quality and of its generalisation characteristics. I n the present paper we propose an original procedure to give a quantit ative description of the uncertainty that derives from the limited siz e of the data set. The effectiveness of the proposed procedure is veri fied by means of experiments. (C) 1998 Published by Elsevier Science B .V. All rights reserved.