Eliminating ATR False Alarms in Complex Underwater Environments with Continuous Learning

AI technologies, such as deep neural networks (DNN), have entered a rapid ascent phase during which they will make significant contributions to the economy and society. It is critical that machine learning is exploited fully for national defense. The easiest deployment of DNNs for the Navy is evaluating growing data sources for intelligence and actionable information. One challenge with this application is that DNNs hallucinate and generate false-positive detections. Even at 99% accuracy, false positives can greatly outnumber true positives if targets occur infrequently. This is best captured by the “mean average precision” (mAP) performance metric. We propose to eliminate false positives with several algorithms that collaborate to boost the mAP of the combined ATR classifier.