Statistical Benchmarks for Neural Network Performance.
Abstract
Much of our effort was devoted to establishing statistical performance bounds for neural networks, acting as pattern classifiers, through improvements to Vapnik-Chervonenkis their (VCT). This theory addresses the interrelationships between the complexity of a network, the amount of training data, and the statistical reliability/performance of the trained networks on independent testing data, and the statistical reliabiiity/performance of the trained network on independent testing data. The troubling chasm between the predictions of VCT and the experience of practitioners remains to be crossed. Our extensive research into reductions of the sample size estimates produced by VCT and into other improvements of the VCT arguments have yet to yield results of practical significance that can serve as advice to neural network designers. In the last year of this program we also initiated work on the use of neural networks as time series forecaster. Our work on time series forecasting has been more successful.
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 31, 1992
- Accession Number
- ADA294937
Entities
People
- Terrence L. Fine
- Thomas W. Parks
Organizations
- Cornell University College of Engineering