Selective cover-based ensemble: Five maybe good enough

Abstract

Generalization capability is a key flag to evaluate the performance of a learning system. Neural network ensemble can greatly improve the generalization capability of a learning system by training many neural networks and composing the result of them. In this paper, based on the theory of neural network ensemble, we present a constructive algorithm to improve the generalization capability of coverage-based neural networks. By construct positive-negative coverage group, the Generalization capability of the CBCNN-based networks can be greatly improved after constructed. Result of the theory analysis and experiments shows that our algorithm can greatly improve the generalization capability even when the initial classification capability of the neural networks is strong.