new mauscript

[tibs@utstat.toronto.edu]

Available in pub/nnboot.ps.Z at utstat.toronto.edu


A comparison of some error estimates for  neural network models

Robert Tibshirani 
Department of Preventive Medicine and Biostatistics
 and
Department of Statistics 
University of Toronto

We discuss a number of methods for estimating the standard error of
predicted values from a neural network (single layer perceptron)
model.  These methods include the delta method based on the Hessian,
bootstrap estimators, and the ``sandwich'' estimator.  The methods are
described and compared in a number of examples.  We find that the
bootstrap methods perform best, partly because they capture variability
due to the  choice of starting weights.


=============================================================
                                            | Rob Tibshirani
  To every man is given the key to          | Dept. of Preventive 
       the gates of heaven;                 |  Medicine and Biostatistics  
      the same key opens the gates of hell. | McMurrich Bldg.
                                            | University of Toronto
           Buddhist proverb                 | Toronto, Canada M5S 1A8
                                
Phone: 1-416-978-4642 (biostats)  Email: tibs at utstat.toronto.edu
         416-978-0673  (stats)
FAX: 1-416-978-8299