Deep Transfer Learning Across Domains, Modalities and Classes

The US Air Force is interested in the ability to transfer learned classifiers from one domain to another. In Phase I, Intelligent Fusion Technology (IFT), Inc. and Rochester Institute of Technology (RIT) developed a robust domain adaptation framework for challenging scenarios where domain shifts are large. In Phase II, the focus is on enhancing the Phase I solution to make it a practical and resilient domain adaptation framework. This will be demonstrated through a practical use case. The Phase II solution will have three areas of focus: open set domain adaptation, model calibration, and continual learning for gradually varying domains. These areas aim to reduce the risk of high-confidence misclassification for military applications and reflect the acquisition of target data in batches over time. The Phase II work will develop a semi-supervised GAN based open set domain adaptation approach that requires only a small number of labeled target domain data to significantly boost the robustness and accuracy of the classification, and a source free unsupervised continue learning for gradually varying domains. Finally, a prototype system with two operating modes will be developed to demonstrate the capability of the proposed solution.