UC Berkeley

Most models in machine learning today are fixed during deployment. As a consequence, a trained model must prepare to be robust to all possible futures, even though only one of them is actually going to happen. The basic idea of test-time training is to train on this future once it arrives in the form of a test instance. Since each test instance arrives without a ground truth label, training is performed with self-supervision. This thesis explores the first steps in realizing this idea, for images, videos and robotics.