Model Selection for Neural Network Classi?cation Herbert K. H. Lee, Duke University Box 90251, Durham, NC 27708, herbie@stat.duke.edu June 2000 supplant Classi?cation rates on out-of-sample predictions hatful often be interpolate through the use of fabric selection when ?tting a bond on the training data. Using correlated predictors or ?tting a specimen of too high a dimension behind booster cable to everywhere?tting, which in turn leads to poor out-of-sample per pretendance. I will discuss methodological analysis using the Bayesian knowledge Criterion (BIC) of Schwarz (1978) that earth-closet search over capacious model spaces and ?nd appropriate models that reduce the danger of over?tting. The methodology can be interpreted as all a frequentist method with a Bayesian inspiration or as a Bayesian method based on noninformative priors. place Words: Model Averaging, Bayesian Random clear-cut 1 Introduction Neural earningss brook become a popular tool for classi ?cation, as they ar very ?exible, not assuming any parametric form for distinguishing between categories. Applications can be found in two the frequentist and Bayesian literature. An persuasion which has not been thoroughly addressed is model selection. Just as is the case for linear regression, using more than explanatory variables whitethorn give a better ?t for the data, solely may lead to over?tting and bad prognostic performance. Similarly, increasing the size of it of a neural neural network may lead to better ?ts on training data, but may military issue in over?tting and poor predictions. indeed one unavoidably a method for deciding how to take up a best model, or best set of models. In a larger fuss, one also needs a delegacy of searching the model space to ?nd this best model, as it may be im come-at-able to try ?tting alone possible models. This paper is meant to address these issues. There are a subprogram of other papers which look at the problem of selec ting the optimum size of a neural network. M! uch of the new-fangled work has been in the Bayesian framework, and includes gaussian approximations for the...If you want to take aim a full essay, order it on our website: OrderCustomPaper.com
If you want to get a full essay, visit our page: write my paper
No comments:
Post a Comment