A Convolutional Fuzzy Min-Max Neural Network for Image Classification

Abstract

Convolutional neural network (CNN) is a well established practice for image classification. In order to learn new classes without forgetting learned ones, CNN models are trained in offline manner which involves re-training of a network considering seen as well as unseen data samples. However, such training takes too much time. This problem is addressed using proposed convolutional fuzzy min-max neural network (CFMNN) avoiding the re-training process. In CFMNN, the online learning ability is added to network by introducing the idea of hyperbox fuzzy sets for CNNs. To evaluate the performance of CFMNN, benchmark datasets such as MNIST, Caltech-101 and CIFAR-100 are used. The experimental results show that drastic reduction in the training time is achieved for online learning of CFMNN. Moreover, compared to existing methods, the proposed CFMNN has compatible or better accuracy.