Sensor Exploitation by Adaptive/Learning Systems (SEALS)

Information-exploitation algorithms are proposed for Air Force sensing and weapons systems, motivated by the inevitable variability and differences seen between conventional training and testing data. Adaptive feedback is proposed, yielding an active-learning framework, wherein the algorithm actively participates in the learning process, by asking questions of the scene under test. The goal of active-learning-based feedback is to efficiently gain insight on the statistics of the testing data, and the relationship of such to the training data. The techniques proposed are based on semi-supervised algorithms. By accounting for the inter-relationships between all of the unlabeled (testing) data, as well as its relationship to the labeled (training) data, semi-supervised algorithms exploit context, providing natural adaptation to changing environments. The Bayesian algorithms yield a statistical measure of confidence in the classification decision, based on the statistical relationship between the training and testing data, and on fundamental limitations of the underlying sensor physics. Rather than simply declaring given items under test as targets or clutter, the proposed algorithms yield a measure of confidence in this declaration. The development of these state-of-the-art algorithms will provide the feedback and adaptively missing in traditional supervised classifiers, significantly advancing the performance of ISR and weapons systems.