conclusions.txt

The idea of test-time training also makes sense for other tasks, such as segmentation and detection, and in other fields, such as speech recognition and natural language process- ing. For machine learning practitioners with prior domain knowledge in their respective fields, their expertise can be leveraged to design better special-purpose self-supervised tasks for test-time training. Researchers for general-purpose self-supervised tasks can also use test-time training as an evaluation benchmark, in addition to the currently prevalent benchmark of pre-training and fine-tuning. More generally, we hope this paper can encourage re- searchers to abandon the self-imposed constraint of a fixed decision boundary for testing, or even the artificial division between training and testing altogether. Our work is but a small step toward a new paradigm where much of the learning happens after a model is deployed. Acknowledgements. This work is supported by NSF grant 1764033, DARPA and Berkeley DeepDrive. This paper took a long time to develop, and benefited from con- versations with many of our colleagues, including Ben Recht and his students Ludwig Schmidt, Vaishaal Shanker and Becca Roelofs; Ravi Teja Mullapudi, Achal Dave and Deva Ramanan; and Armin Askari, Allan Jabri, Ashish Kumar, Angjoo Kanazawa and Jitendra Malik. Test-Time Training with Self-Supervision for Generalization under Distribution Shifts