Dissertation Talk: Learning to generalize in dynamic environments: Fully test-time adaptation by online and offline optimization

  • Dequan Wang, UC Berkeley
A model must adapt itself to generalize to new and different data during testing. In this setting of fully test-time adaptation the model has only the test data and its own parameters. We propose to adapt by test entropy minimization (tent): we optimize the model for confidence as measured by the entropy of its predictions. Tent reduces generalization error for image classification on corrupted...