Human-Machine Teaming with Machine Learning Algorithms

Characterizing and understanding the interactions between human and machine plays an important role in extracting the most out of our machine learning algorithms while reducing human workload. We propose to develop a software prototype system that reduces user workload of exploiting AI algorithms for imagery exploitation. We will design a user-friendly system for content matching with interactive query refinement (IQR) and conduct human-machine teaming (HMT) user studies to inform the user interface and algorithm design choices. On Phase I. we have produced protocols for the HMT user studies based on reducing user workload through IQR. The feasibility study showed that the proposed novel approach of using active learning with IQR greatly reduces the user workload in searching FMV archives. We propose to develop an enhanced prototype based on such a design and further add novel techniques to improve product quality with reduced user workload. We propose to use techniques like knowledge distillation to avoid retraining deep learning algorithms. In order to establish user trust in AI, new cluster-based explanations with faster saliency map will be developed. The proposed technology should also generalize to other data sources like mid and high-altitude collection platforms and publicly available information.