Unified granular neural networks for pattern classification

Motivation in the development of granular neural network (GNN) with the principle of granular computing (GrC) is to obtain possible degree of transparency in the network architecture and its operational steps, which are bottleneck for conventional neural network (NN). In addition, GNN provides improved performance and at the same time poses less computational burden compared to conventional NN and fuzzy NN (FNN). Topologically, the nodes of different layers of FNNs are fully connected whereas it is partial in case of GNNs. Further, the architectures of GNN is determined based on the extracted rules from the domain information. However, selection of relevant rules for GNN is a tedious task. To mitigate this, the present paper aims to develop a pattern classification model in the framework of unification of GNNs and tries to avoid the searching of optimum number of rules and the best combination of rules. The unified model thus works with a set of different rules-based GNNs and derives the final decision from the individual GNN. Each of these GNNs takes the class-supportive fuzzy granulated features of the input in order to preserve the feature-wise belonging information to different classes. These granulated features provide improved class discriminatory information for the classification of data sets with ill-defined and overlapping class boundaries. The proposed model thus explores mutually the advantages of class-supportive fuzzy granulation of features, informative rules-based GNNs and unification of GNNs. Superiority of the model to similar other methods are justified in terms of various performance measurement indexes using different benchmark data sets including remote sensing imagery.