Article ID | Journal | Published Year | Pages | File Type |
---|---|---|---|---|
9650506 | Engineering Applications of Artificial Intelligence | 2005 | 11 Pages |
Abstract
This paper proposes a framework for the training of learning systems for regression when (i) the number of examples is small and contains interdependencies, and (ii) each sample consists of large quantities of discrete data that are functional in nature. The objective is to achieve robust yet nonlinear relations between inputs and outputs. In this study, laser scans of the trunk surface and reconstructions of spinal data from X-rays from scoliosis patients were functionally represented as surfaces and curves. Leading functional principal component coefficients thereof constituted comprehensive features, and achieved sufficient dimensionality reduction for the prediction of spine from trunk. As a learning method, support vector regression (SVR) was chosen for its strong generalizability capability that stems from penalizing model complexity. A first robust prediction in this research application was obtained, with coefficients of determination for leading outputs of 0.70 and 0.82, respectively, in the test set. Those translated to a spinal curve prediction L2-error of 3.61Â mm, comparable to measurement error in data.
Keywords
Related Topics
Physical Sciences and Engineering
Computer Science
Artificial Intelligence
Authors
Charles Bergeron, Farida Cheriet, Janet Ronsky, Ronald Zernicke, Hubert Labelle,