NONLINEAR FUNCTION MODELING WITH NEURAL NETS.
Abstract
Some of the general properties of linear neural nets are reviewed and their applications to adaptive decorrelation of a set of correlated signals and to adaptive inversion of matrices are outlined. The effects on modeling accuracy using non-linear preprocessing by quantizing the inputs to adaptive nets have been studied. The properties of nonlinear modeling nets with and without linear pole-adaption are discussed and the information and convergence aspects of the several types of modelers are considered. A powerful quantizer for high accuracy modeling is derived. It transmits in separable forms, and without cross-coupling, information specifying the class to which the input belongs and also the value of the input at the instant. This technique affords piecewise-linear approximation of any single-valued nonlinear transformation of the input; it also yields rapid convergence of the adaptive weights and high accuracy of modeling. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 14, 1967
- Accession Number
- AD0816224
Entities
People
- Frederic D. Powell
- Johannes G. Goerner
Organizations
- Bell Aircraft Corporation