Student Seminar Series - October 5, 2004
University of Minnesota
School of Statistics
College of Liberal Arts

Prediction Intervals for Neural Networks: Literature Review and Future Directions

Sriram Thirumalai


Tuesday, October 5, 2004
1:00 PM, B60 Ford Hall
Minneapolis, East Bank Campus

Refreshments at 12:30 PM
300 Ford Hall

Abstract

 Neural Networks are statistical models that model the response as a nonlinear function of various linear combinations of the predictors. The multi-layer non-linear functional form of neural networks that is flexibly complex and yet tractable, its relatively distribution-free approach and its ability to deal with large data sets makes it a valuable alternative for statistical analysis in a host of applications. A critical part of the use of neural networks in statistical predictions is obtaining the confidence intervals about the mean and future predictions. This project first reviews two broad approaches from the extant literature to estimate prediction intervals for neural networks. Next, building on the extant literature it suggests alternate approaches to calculate the prediction intervals.