Novel supervised fuzzy adaptive resonance theory (SF-ART) neural network for pattern recognition

Abstract

This paper introduces a novel two-stage artificial neural network so-called supervised fuzzy adaptive resonance theory (SF-ART) which is utilized for training as well as for recall. First stage, pre-classification level, includes an unsupervised neural network called fuzzy adaptive resonance theory (K-ART) to roughly classify input data to M classes. M is variable and it depends to input features scattering and learning parameters of F-ART. The second stage, post (affine)-classification level, only includes an array with M cells called affine look-up table (ALT), By representing input data to the F-ART (first stage), one of F-ART output neurons are wined. The index of this wined output is used as reference address to call and upload output desired value in the related cell of ALT in training mode (supervised learning), or to recall and introduce its value as the final decision (wined class) in testing mode. The SF-ART is evaluated and compared with existing supervised neural networks on a variety of some well-known pattern classification problems.