Neuron Requirements for Classification
Abstract
The feed forward layered neural networks holds great promise for application to classification problems. Determination of the sizes of the layers is an important network design problem. This report treats the neuron requirement question from the geometric viewpoint. Threshold neurons correspond to cutting planes in the Euclidean space of input patterns. Bounds on the minimum number of first-layer neurons are determined as functions of the partition sizes of the training data sets. Bounds are also proved for convex pattern classes. Measures of separability of the training data are defined in order to emphasize the dependence of the design parameters upon the geometry of the classes.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 1991
- Accession Number
- ADA238003
Entities
People
- W. O. Alltop
Organizations
- Naval Air Weapons Station China Lake