A geometric method to obtain error-correcting classification by neural networks with fewer hidden units
Věra Kůrková, Paul C. Kainen · 2002
Coding the output of a neural network can reduce the output dimension but may also increase the requirements for approximation accuracy tolerated by error-correcting. Using a generalization of orthogonality, we propose a method for error-correcting classification by feedforward neural networks that reduces the number of hidden units for a variety of classes of patterns.