Class-Specific Fuzzy Hypersphere Neural Network
A.B. Kulkarni, Sanjiv Vedu Bonde, Uday V. Kulkarni · Procedia Computer Science · 2018
A new Fuzzy Hypersphere Neural Network (FHNN) classifier is proposed in the framework of Radial Basis Function Neural Network (RBFNN) for which two class specific fuzzy clustering algorithms are suggested for the creation of fuzzy set hyperspheres (FSHs) in the hidden layer and the traditional least mean square (LMS) algorithm used to determine the weights between hidden to output layer is avoided by assigning the appropriate binary weights during training. The FHNN classifier utilizes FSH as pattern cluster and classes are represented by union of FSHs. The proposed fuzzy clustering algorithms are developed on the basis of inter-class and intra-class fuzzy membership metrics with maximum coverage of data points to create the FSHs in the hidden layer of FHNN. The weights between hidden to output layer of FHNN are adapted simultaneously during the creation of FSHs in the hidden layer. Unlike RBFNN and other similar models, the learning is: fast, independent of tuning parameters, formulates precise centroids and widths or radii of the FSHs, leading to 100% training accuracy on any training set. The learning rule 1 proposed for the FHNN is insusceptible to the order of data presentation and always results in identical clusters showing good data visualization, while the learning rule 2 eradicates the overlap between inter-class FSHs. The performance of FHNN is compared with the RBFNNs using twelve benchmark data sets. The empirical findings demonstrate that the FHNN is highly efficient classifier.