CEATR (Common Embeddings for Aided and automatic Target Recognition)

To support the U.S. Army's needs for approaches to training deep-learning (DL) algorithms with scarce visible/IR data and data annotations, ARiA proposes CEATR (Common Embeddings for Aided and automatic Target Recognition), a learning framework designed to learn robust embeddings to reduce dataset shift--that is, the variations in datasets that prevent models from generalizing.  ARiA will build on our development of DL for transfer learning from synthetic-to-measured and measured-to-measured imagery to demonstrate the feasiblity of CEATR to: (1) learn a robust feature space that is invariant to differences in datasets caused by sensor or data-acquisition variations, (2) learns to transfer features across a variety of synthetic and measured data sources, (3) generalizes to novel imagery and targets with scarce labels and annotations. Phase I effort will (1) research, document, and publish techniques and design of a learning framework, utilizing deep neural networks and statistical-learning techniques, that makes use of multiple data sources to learn a common embedding space that is robust to dataset shift; (2) demonstrate that CEATR can feasible meet U.S. Army needs by developing a methodology to reduce post mission analysis and manual annotations for training algorithms on novel visible/IR imagery; and (3) develop conditions under which the proposed technology can be applied and establish that CEATR can be developed into a useful product for the U.S. Army that is compatible with existing decision chains and workflows across multiple U.S. Army, Department of Defense (DoD) and commercial systems.