In situ learning techniques for underwater object classification.

Abstract

A Bayesian in situ learning framework is presented that integrates several techniques to dynamically learn and adapt to variable sensing environments and the resulting manifestations of objects of interest. Sensing data are often collected from an environment for which little or no a prior knowledge is available. The presented framework exploits the contextual information of unlabeled data by learning a semi‐supervised classifier on the complete data manifold. An active learning component augments limited labeled data through the adaptive selection of unlabeled samples that, given acquired labels, minimize uncertainty in the classifier. Both myopic and nonmyopic active learning methods are presented, where nonmyopic selection of the most‐informative subset of samples is an extension of the myopic approach and leverages properties of submodular set functions. Additionally, the framework addresses the problem of imperfect acquired labels by accounting for uncertainty in a label within the classifier design. The benefits of the in situ learning techniques are demonstrated through the application to underwater object recognition. [This work was supported by the Office of Naval Research.]