Distributed Deep Learning and Sensing

Compared to current practices of today’s processing, exploitation, and dissemination (PED) cycles for fusing the outputs of single-sensor-processing chains during the feature- or decision-level fusion tasks, combining modalities at the data level to learn the target-signatures offers an opportunity to utilize the mutual information among the raw signals generated by the same target. Hence, the automated decision aids which exploit the correlations across signals and sensors could potentially offer a significant improvement over the current practice, leading to increased probability of correct target identification/classification decisions while reducing the concomitant false-alarm rates. To harness the power of Deep Learning algorithms for streamlining the distributed (i.e., multi-platform), multi-modal/multi-source, real-time/on-the-fly signal processing, feature learning, and concomitant generation of fused intelligence for target and/or event detection/classification, we offer to develop KARMA (Knowledge-to-Action with Remote-and-networked Machine-learning Algorithms), an engineering toolbox that facilitates the optimization and testing of distributed Deep Learning architectures and algorithms under varying ISR scenarios, platform lay-down geometry configurations, cross-platform communication bandwidth constraints, and sensor resolutions and target-of-interest types. KARMA will lead to superior distributed sensing techniques employing deep learning networks that span disparate sensors and widely separated platforms to improve detection of unanticipated threats, events, and targets.