A General Reflex Fuzzy Min-Max Neural Network

A General Reflex Fuzzy Min-Max Neural Network" (GRFMN) is presented. GRFMN is capable to extract the underlying structure of the data by means of supervised, unsupervised and partially supervised learning. Learning under partial supervision is of high importance for the practical implementation of pattern recognition systems, as it may not be always feasible to get a fully labeled dataset for training or cost of labeling all samples is not affordable. GRFMN applies the acquired knowledge to learn a mixture of labeled and unlabeled data. It uses aggregation of hyperbox fuzzy sets to represent a class or cluster. A novel reflex mechanism inspired from human brain is used to solve the problem of class overlaps. The reflex mechanism consists of compensatory neurons which become active if the test data belongs to an overlap region of two or more hyperboxes representing different classes. These neurons help to approximate the complex topology of data in a better way. The proposed new learning approach to deal with partially labeled data and inclusion of compensatory neurons, has improved the performance of GRFMN significantly. The advantages of GRFMN are its ability to learn the data in a single pass through and no requirement of retraining while adding a new class or deleting an existing class. The performance of GRFMN is compared with "General Fuzzy Min-Max Neural Network" proposed by Gabrys and Bargiela. The experimental results on real datasets show a better performance of GRFMN.