Optimize Recognition Algorithms and Classifiers by Learning to Explore (ORACLE)

Commercial imaging satellites capture the majority of the planet’s surface every day, generating millions of images. Searching such data for specific objects of interest is extremely difficult to do manually without automated computer image analytics. NGA is seeking a novel scientific formalism and tools for training overhead image detectors on rare and new objects of interest, for which large, annotated datasets are not available. Aptima proposes to develop a formal mathematical framework and supporting learning and planning analytics that Optimize Recognition Algorithms and Classifiers by Learning to Explore (ORACLE). ORACLE will extract low-dimensional content encoding from the overhead images to generalize how the detector and labeler produce their decision outcomes, identify what factors affect the evolution of the state of the detector, and use this information to optimize the bootstrapping policy. ORACLE will minimize the same objective function, called variational free energy, which is a formal bound on uncertainty, for both learning (representation model) and planning (selection of images for bootstrapping). It will allow ORACLE to estimate the impact of images on detectors via (1) the value of detection goals (measuring how well the detector can find objects of interest), and (2) the value of information gain (resolving internal uncertainty).