Test-Time Training with Masked Autoencoders

09/15/2022

∙

by   Yossi Gandelsman, et al.

∙

4

∙

share

Test-time training adapts to a new test distribution on the fly by optimizing a model for each test input using self-supervision. In this paper, we use masked autoencoders for this one-sample learning problem. Empirically, our simple method improves generalization on many visual benchmarks for distribution shifts. Theoretically, we characterize this improvement in terms of the bias-variance trade-off.