Test-Time Training with Masked Autoencoders

Yossi Gandelsman, Yu Sun, Xinlei Chen, Alexei A Efros

Keywords: Computer Vision, Test-Time Training, Masked Auto-Encoder

TL;DR: We show how applying masked autoencoding to train on each unlabeled test sample before making a prediction improves generalization.

Abstract:

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.

Community Implementations: [ 1 code implementation](https://www.catalyzex.com/paper/arxiv:2209.07522/code)