Experimental analysis of generalization capability based on information criteria
Takashi Onoda · 2002
Feed-forward neural networks are nonlinear parametric models that can approximate any continuous input-output relation. The quality of the approximation can be measured by the generalization capability. This paper presents an experimental analysis to measure the quality of the approximation or the generalization capability for artificial feed-forward neural networks. Our approach analyzes the relation between some information criteria and the generalization capability statistically and experimentally. These information criteria are used to select the best model and the generalization capability is a measure of the quality of approximation. We make clear the feature of some information criteria; Akaike's criterion, the network criterion, the information criterion of numbers of learning examples and of independently adjusted parameters, and neural net criterion. Moreover, our experiments compare the generalization capability with the value of these information criteria. Finally, this paper shows that neural net criterion is effective to measure the quality of the approximation or the generalization capability for artificial feed-forward neural networks.