IMPROVING GENERALIZATION PERFORMANCE IN CNN CLASSIFIER USING A MLP-BASED TECHNIQUE

Abstract

This paper presents a multi-layer perceptron (MLP)-based technique for improving generalization performance in condensed nearest-neighbor (CNN) classifier. The CNN classifier is simple and efficient in time due to its condensed set of prototypes. However, its generalization performance is not as good as that of nearest-neighbor (NN) classifier that uses the complete large training data set. The developed MLP-based technique is used to modify the condensed set of prototypes of CNN classifier in order to generate or enhance the useful features of such prototypes so that the CNN classifier could also achieve good generalization performance. The improving in the performance of the developed CNN classifier is shown by experimental results.