A feed-forward neural architecture for estimating the fuzziness in non-sparse data sets

This paper introduces a neural architecture capable of estimating the fuzzy membership functions of nonsparse or overlapping classes of a data set. This architecture consists of m second-order feedforward neural networks, one for each of the m classes in the data set. The response of the ith second-order feed-forward network for any input feature vector gives the membership value of this feature vector to the ith class. The weight matrices of these second-order feed-forward networks are provided by m single-layered feed-forward neural networks. The inputs of these networks are formed by the feature vectors belonging to different classes and their weights are updated using generalized Hebbian learning rules. This architecture can also be used as a feed-forward neural classifier for nonsparse data sets. The IRIS data set and an artificial data set are used to test the performance of the proposed architecture. Experimental results show that the training process can be completed in real time