name |
email |
phone |
|
Walter D Bennette |
walter.bennette.1@us.af.mil |
315-330-4957 |
The need for increased levels of autonomy has significantly risen within the Air Force. Thus, machine learning tools that enable intelligent systems have become essential. However, analysts and operators are often reluctant to adopt these tools due to a lack of understanding – treating machine learning as a black box that introduces significant mission risk. Although one may hope that improving machine learning performance would address this issue, there is in fact a trade-off: increased effectiveness often comes at the cost of increased complexity. Increased complexity then leads to a lack of transparency in understanding machine learning methods. In particular, it becomes unclear when such methods will succeed or fail, and why they will fail. This limits the adoption of intelligent systems. This topic focuses on building trust in machine learning models by designing models that fail elegantly. Of particular interest are model calibration techniques for object detection and classification, novelty detection, open-set recognition, and post-hoc filters to identify instances prone to causing model failure. Other topics related to this area will also be considered.
Find and choose an agency to see details and to explore individual opportunities.