The Influence of Network Structure on the Performance of Learning Systems

摘要

The report summarizes research on means for improving the discrimination capability of linear single-gain layer networks. Improvements are considered via (a) nonlinear pre-processors, such as the A/D converter, the quantizer, and other encoders, the generation of functionals of the individual pattern components, and (b) cascading of these networks. All result essentially in introducing a nonlinear decision surface instead of the decision hyperplane as defined by a linear net. Significant improvements are obtainable that widen the application of such networks to practical problems. As an example, the modeling of an air cushion vehicle is discussed. (Author)