Learning dimensionality-reduced classifiers for information fusion

Abstract

The fusion of multimodal sensor information often requires learning decision rules from samples of high-dimensional data. Each data dimension may only be weakly informative for the detection problem of interest. Also, it is not known a priori which components combine to form a lower-dimensional feature space that is most informative. To learn both the combination of dimensions and the decision rule specified in the reduced-dimensional space together, we jointly optimize the linear dimensionality reduction and margin-based supervised classification problems, representing dimensionality reduction by matrices on the Stiefel manifold. We describe how the learning procedure and resulting decision rule can be implemented in parallel, serial, and tree-structured fusion networks.