Sensitivity Analysis for Neural Networks by Daniel S. Yeung Ian Cloete Daming Shi & Wing W. Y. Ng

Author:Daniel S. Yeung, Ian Cloete, Daming Shi & Wing W. Y. Ng
Language: eng
Format: epub
Publisher: Springer Berlin Heidelberg, Berlin, Heidelberg

5.3 Architecture Selection Using the Error Bound

The selection of the number of hidden neurons in the RBFNN is usually done by Sequential Learning (Huang et al., 2005) or by ad hoc choice. The Sequential Learning technique only makes use of the training error to determine the number of hidden neurons, without any reference to the generalization capability. Moreover, Huang et al. (2005) and Liang et al. (2006) assume the classifier does not have prior knowledge about the number of training samples while Kaminski and Strumillo (1997) and Gomm and Yu (2000) assume that it does. For ease of comparison with other architecture selection methods, we assume that the number of training samples in our experiments is known to the classifier. In this section, we describe a new technique based on to find the optimal number of hidden neurons that makes use of the generalization capability of the RBFNN.

For any given threshold a on the generalization error bound (), the localized generalization error model allows us to find the best classifier by maximizing Q, assuming that the MSE of all samples within the Q-Union is smaller than a. One can formulate the architecture selection problem as a Maximal Coverage Classification problem with Selected Generalization error bound (MC2SG), i.e.,

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.