Classification-Oriented Neural Networks with Granular Processing

This paper presents a classification framework of granular neural networks (GNNs) designed for classification tasks. Built upon conventional neural network architectures, the framework introduces interval-valued (granular) connections, where each interval forms an information granule around a numeric weight. This design is proposed to utilize a granular approach in classification tasks. This representation provides a quantifiable measure of uncertainty and interpretability in classification decisions. By optimally distributing information granularity across network connections, the proposed model seeks to maximize classification confidence while preserving specificity. To achieve this objective, a single-objective particle swarm optimization (PSO) algorithm is employed, guided by classification coverage and specificity criteria. The proposed approach is validated using benchmark datasets and evaluated through multiple performance indices derived from the Interval Neutrosophic Set (INS) and granular computing domains.