Nonlinear principal predictor analysis using neural networks

Principal predictor analysis is a linear technique which fits between regression and canonical correlation analysis in terms of the complexity of its architecture. This study introduces a new neural network approach for performing nonlinear principal predictor analysis. The utility of this approach is demonstrated via two test problems. The first, using synthetic data, gauges the ability of the model to extract known modes of variability from datasets with increasing noise levels. The second, based on the Lorenz system of equations, considers performance in the context of nonlinear prediction. Results suggest that nonlinear principal predictor analysis performs better than nonlinear canonical correlation analysis. In addition, nonlinear principal predictor modes may be extracted in less time than modes from nonlinear canonical correlation analysis.