IRIS-HEP Fellow: Hector Castro Noguez
Fellowship dates: Jun – Aug, 2020
Home Institution: Boston University
Project: Domain Adaptation via Histogram Loss ComponentWhile gradient reversal layers are shown to be successful, it is possible to take a more direct approach to improving the agreement between data and simulation. Rather than discourage the DNN from learning features which allow it to distinguish between examples from the source and target domain (as done for the gradient reversal layer), we propose to explicitly reward the DNN for minimizing differences between distributions in the source and target domains.
More information: My project proposal
Samuel May (Boston University)
Indara Suarez (Boston University)
- 27 Oct 2021 - "Domain Adaptation via Histogram Loss Component", Hector Castro Noguez, IRIS-HEP Topical Meetings Recording: Domain Adaptation via Histogram Loss Component