Designing Kernel Scheme for Classifiers Fusion

Abstract

In this paper, we propose a special fusion<br><br>method for combining ensembles of base classifiers<br><br>utilizing new neural networks in order to improve<br><br>overall efficiency of classification. While ensembles are<br><br>designed such that each classifier is trained<br><br>independently while the decision fusion is performed as<br><br>a final procedure, in this method, we would be<br><br>interested in making the fusion process more adaptive<br><br>and efficient.<br><br>This new combiner, called Neural Network Kernel<br><br>Least Mean Square1, attempts to fuse outputs of the<br><br>ensembles of classifiers. The proposed Neural Network<br><br>has some special properties such as Kernel abilities,<br><br>Least Mean Square features, easy learning over<br><br>variants of patterns and traditional neuron capabilities.<br><br>Neural Network Kernel Least Mean Square is a special<br><br>neuron which is trained with Kernel Least Mean<br><br>Square properties. This new neuron is used as a<br><br>classifiers combiner to fuse outputs of base neural<br><br>network classifiers. Performance of this method is<br><br>analyzed and compared with other fusion methods. The<br><br>analysis represents higher performance of our new<br><br>method as opposed to others.