SoTTA: Robust Test-Time Adaptation on Noisy Data Streams

Taesik Gong,Yewon Kim,Taeckyung Lee,Sorn Chottananurak,Sung-Ju Lee

2023-10-16

Abstract:Test-time adaptation (TTA) aims to address distributional shifts between training and testing data using only unlabeled test data streams for continual model adaptation. However, most TTA methods assume benign test streams, while test samples could be unexpectedly diverse in the wild. For instance, an unseen object or noise could appear in autonomous driving. This leads to a new threat to existing TTA algorithms; we found that prior TTA algorithms suffer from those noisy test samples as they blindly adapt to incoming samples. To address this problem, we present Screening-out Test-Time Adaptation (SoTTA), a novel TTA algorithm that is robust to noisy samples. The key enabler of SoTTA is two-fold: (i) input-wise robustness via high-confidence uniform-class sampling that effectively filters out the impact of noisy samples and (ii) parameter-wise robustness via entropy-sharpness minimization that improves the robustness of model parameters against large gradients from noisy samples. Our evaluation with standard TTA benchmarks with various noisy scenarios shows that our method outperforms state-of-the-art TTA methods under the presence of noisy samples and achieves comparable accuracy to those methods without noisy samples. The source code is available at <a class="link-external link-https" href="https://github.com/taeckyung/SoTTA" rel="external noopener nofollow">this https URL</a> .

Machine Learning

What problem does this paper attempt to address?

The paper attempts to address the issue of how to handle the impact of noisy samples on model performance during Test-Time Adaptation (TTA). Existing TTA methods typically assume that the test data stream contains only benign samples, overlooking the fact that in real-world scenarios, test data may include various noisy samples. For example, in autonomous driving scenarios, cameras may be affected by dust, strong light, and other interferences, leading to a decline in model performance.

The authors found that when noisy samples are mixed into the test data, the performance of existing TTA algorithms significantly decreases. To solve this problem, the paper proposes a new TTA algorithm—SoTTA (Screening-out Test-Time Adaptation), which aims to improve the robustness of the model in the presence of noisy samples.

Specifically, SoTTA achieves robustness to noisy samples through the following two aspects:

1. **Input-level robustness**: By filtering out noisy samples through High-confidence Uniform-class Sampling (HUS), ensuring that the model is updated only with high-quality samples.
2. **Parameter-level robustness**: By using Entropy-sharpness Minimization (ESM) to make the model parameters more resistant to gradient perturbations caused by noisy samples.

Through these methods, SoTTA is able to maintain high classification accuracy in various noisy scenarios, significantly outperforming existing TTA algorithms.