Virtual Reality for Multi-INT Deep Learning (VR-MDL)

Recent advances and successes of deep learning neural networks (DLNN) techniques and architectures have been well publicized over the last several years. Voluminous, high-quality and annotated training data, or trial and error in a realistic environment, is required to achieve the promised performance potential of DLNNs. Unfortunately for DoD and/or Intelligence Community (IC) applications of multi-INT fusion, there is a dearth of high-quality, annotated training data covering all contingencies in highly-contested adversarial environments. And there is really no conceivable practical way in the future to obtain such data. Although we routinely collect a “fire house” of multi-INT data, it is not suitable for DLNN training since it is not processed, vetted, and annotated. Not to mention that true targets of interest are embedded in enormous amounts of clutter, noise and other extraneous signals. In this project, ISL and Ohio University propose a novel approach for creating the requisite training data and environment by combining state-of-the-art DLNN methods with a high-fidelity, multi-physics-based modeling and simulation framework for multi-INT sensor systems.