Learning to Adapt to Online Streams with Distribution Shifts

TL;DR

This paper introduces a meta-learning approach enabling models to adapt continually to online data streams with distribution shifts, improving performance without relying on large test batches.

Contribution

We propose a novel meta-learning method for test-time adaptation that handles online streams with distribution shifts, overcoming batch size limitations and modeling the distribution evolution.

Findings

Consistent improvements over state-of-the-art methods on benchmark datasets.

Effective adaptation to distribution shifts in online streaming scenarios.

Superior performance in video segmentation tasks.

Abstract

Test-time adaptation (TTA) is a technique used to reduce distribution gaps between the training and testing sets by leveraging unlabeled test data during inference. In this work, we expand TTA to a more practical scenario, where the test data comes in the form of online streams that experience distribution shifts over time. Existing approaches face two challenges: reliance on a large test data batch from the same domain and the absence of explicitly modeling the continual distribution evolution process. To address both challenges, we propose a meta-learning approach that teaches the network to adapt to distribution-shifting online streams during meta-training. As a result, the trained model can perform continual adaptation to distribution shifts in testing, regardless of the batch size restriction, as it has learned during training. We conducted extensive experiments on benchmarking datasets for TTA, incorporating a broad range of online distribution-shifting settings. Our results showed consistent improvements over state-of-the-art methods, indicating the effectiveness of our approach. In addition, we achieved superior performance in the video segmentation task, highlighting the potential of our method for real-world applications.