Technical Report Number
Computing Methodologies, Computer Applications, Theory of Computation
Constructive learning algorithms offer an attractive approach for incremental construction of potentially near-minimal neural network architectures for pattern classification tasks. These algorithms help overcome the need for ad-hoc and often inappropriate choice of network topology in the use of algorithms that search for a suitable weight setting in an a-priori fixed network architecture. Several such algorithms proposed in the literature have been shown to converge to zero classification errors (under certain assumptions) on finite, non-contradictory training sets in 2-category classification tasks. The convergence proofs for each of these algorithms (with the exception of the Upstart and the Perceptron Cascade) rely on the assumption that the pattern attributes are either binary or bipolar valued. This paper explores multi-category extensions of several constructive neural network learning algorithms for classification tasks where the input patterns may take on real-valued attributes. In each case, we establish the convergence to zero classification errors on a multi-category classification task. Results of experiments with non-linearly separable multi-category datasets demonstrate the feasibility of this approach and suggest several interesting directions for future research.