A Comparison of Methods for Learning of Highly Non-separable Problems
ICAISC '08 Proceedings of the 9th international conference on Artificial Intelligence and Soft Computing
| Hi-index | 0.00 |
A fast iterative algorithm is proposed for the construction and the learning of a neural net achieving a classification task, with an input layer, one intermediate layer, and an output layer. The network is able to learn an arbitrary training set. The algorithm does not depend on a special learning scheme (e.g., the couplings can be determined by modified Hebbian prescriptions or by more complex learning procedures). During the process the intermediate units are constructed systematically by collecting the patterns into smaller subsets. For simplicity, we consider only the case of one output neuron, but actually this restriction is not necessary.