A Reflex Fuzzy Min Max Neural Network for Semi-supervised Learning

摘要

"A Reflex Fuzzy Min-Max Neural network" (RFMN) based on hyperbox fuzzy sets is discussed. RFMN is capable to extract the underlying structure of the data with supervised, unsupervised, and semi-supervised learning techniques. Semi-supervised learning is of high importance for the practical implementation of pattern recognition systems. RFMN can handle point and granular data simultaneously. Its architecture consists of a reflex mechanism inspired from human brain, and is implemented using concept of compensatory neurons. An important stage in the training phase is to manage class overlaps. Compensatory neurons are trained to handle these overlaps and are helpful to approximate the complex topology of data in a better way. The proposed new semi-supervised learning approach has improved the accuracy of RFMN significantly. Moreover, RFMN performance is almost independent of the maximum hyperbox size. The advantage of RFMN is that it can learn data in a single pass through (on-line). RFMN performance is compared with "General Fuzzy Min-max Neural network" proposed by Gabrys and Bargiela on several datasets.