Stopped Training and Other Remedies for Overfitting

Stopped Training and Other Remedies for Overfitting (1995)

Cached

Download Links

[mars.elcom.nitech.ac.jp]

by Warren S. Sarle

Venue:	Proceedings of the 27th Symposium on the Interface of Computing Science and Statistics
Citations:	40 - 0 self

Versions

Version History

Metadata Version 3

User correction supplied by SystemCorrections

Datum	Value	Source
TITLE	Stopped Training and Other Remedies for Overfitting	user correction - Legacy Corrections
AUTHOR NAME	Warren S. Sarle	SVM HeaderParse 0.1
AUTHOR AFFIL	SAS Institute Inc	SVM HeaderParse 0.2
AUTHOR ADDR	, SAS Campus Drive, Cary, NC 27513, USA	SVM HeaderParse 0.1
ABSTRACT	Overfitting is the most serious problem in neural network training; stopped training is the most common solution to this problem. Yet the neural network literature contains little systematic investigation of the properties of stopped training and no comparisons with statistical methods of dealing with overfitting. This paper presents the results of simulations designed to investigate these issues. Introduction The most important unresolved problem in practical applications of feedforward neural networks (NNs) is how to determine the appropriate complexity of the network. This problem is similar to the problem of choosing the degree of smoothing in nonparametric estimation. But overfitting in NNs can have much more serious consequences than can undersmoothing in most nonparametric estimation methods. NNs, like high-order polynomials, can produce wild predictions far beyond the range of the data. There are two main approaches to controlling the complexity of a NN: model selection and re...	user correction - Legacy Corrections
YEAR	1995	INFERENCE
VENUE	Proceedings of the 27th Symposium on the Interface of Computing Science and Statistics	INFERENCE
VENUE TYPE	CONFERENCE	INFERENCE
PAGES	352--360	INFERENCE
CITATIONS	10 found	ParsCit 1.0