Intelligent Automation, Inc. — Department of Defense SBIR Phase I: NGA203-005

Description

Detecting relevant objects of interest in large datasets using artificial intelligence techniques is very appealing. However, most state of the art approaches use deep neural network techniques -- requiring millions of human annotated training examples. These datasets mostly come from academia and don’t transfer well to novel domains. Even with the use of smaller annotated datasets and an application of transfer learning techniques, human annotations are required, which can be impossible to obtain in limited data scenarios or when annotation time from analysts is not available. Additionally, the end user of these technologies has no way of identifying why the system made its decisions and is unable to determine how to fix the errors. A system designed to address these problems should (i) learn how to detect new objects of interest quickly; (ii) be able to detect across domains and camera view-points; and (iii) be more accessible and human understandable to an end user who is not a machine learning expert. Intelligent Automation, Inc. (IAI) propose to develop a novel mathematical model and training paradigm to build a system that creates automated annotation analytics from limited or zero training samples from new datasets and domains to build robust annotation analytics.