INVESTIGATING FEEDFORWARD NEURAL NETWORKS WITH RESPECT TO THE REJECTION OF SPURIOUS PATTERNS

The reliability of feedforward neural networks with respect to the rej ection of patterns not belonging to the defined training classes is in vestigated. It is shown how networks with different activation functio ns and propagation rules construct the decision regions in the pattern space and, therefore, affect the network's performance in dealing wit h spurious information. A modification to the standard MLP structure i s described to enhance its reliability in this respect.