Dr. Walter Bennette of Air Force Research Laboratory, Information Directorate

Traditional classification and recognition models are trained with an assumption that all possible classes are encountered during model training. When such models operate in the open-world, they can encounter observations from previously unknown classes, and are prone to making erroneous high confidence predictions. In this work, we define the problem of hierarchical Open-Set Recognition (OSR) which seeks classification models that are able to identify if an observation belongs to an unknown class, and to provide information about how the unknown class relates to the known classes. As part of this work we introduce three novel approaches to  hierarchical OSR, and a novel evaluation metric that conveys the ability of hierarchical OSR models to provide partial information for predictions. Through empirical results we demonstrate that one of our novel hierarchical OSR approaches outperforms competing methods. Finally, we introduce a use case that shows the benefit of hierarchical OSR for Automatic Target Recognition applications.

Biography:

Dr. Walter Bennette is a Senior Research Engineer at the Air Force Research Laboratory, Information Directorate.  His research interests lie in the intersection between operations research and machine learning.  Recent works have focused on the efficient evaluation of machine learning models, and the creation of machine learning models for open world settings.  Currently, Dr. Bennette is serving as the Special Assistant to the Air Force's Chief Scientist.

Email: