Test-Time Training with Masked Autoencoders
This addresses the challenge of distribution shifts in visual tasks, but it appears incremental as it applies an existing method (masked autoencoders) to a known problem.
The paper tackles the problem of adapting models to new test distributions using test-time training with masked autoencoders, resulting in improved generalization on visual benchmarks for distribution shifts, though no specific numbers are provided.
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.