Improving generalization in backpropagation networks architectures

This paper gives a prototype recognizer that uses rough reduction module to find the optimal representation for backpropagation networks. The proposed approach exhibits a hybrid methodology for feedforward neural networks and rough set theory. The system is a two stand alone subsystems, in which the output of the first is fed to the second for recognition tasks. The system is investigated for detection and recognition of patterns present in an image. The rough module deals with uncertainty and irrelevant observations inherited in the data. The novel architecture integrates the two approaches to recognize pattern efficiently, with minimal neurons architecture