DATTA: Towards Diversity Adaptive Test-Time Adaptation in Dynamic Wild World

Chuyang Ye,Dongyan Wei,Zhendong Liu,Yuanyi Pang,Yixi Lin,Jiarong Liao,Qinting Jiang,Xianghua Fu,Qing Li,Jingyan Jiang

2024-08-15

Abstract:Test-time adaptation (TTA) effectively addresses distribution shifts between training and testing data by adjusting models on test samples, which is crucial for improving model inference in real-world applications. However, traditional TTA methods typically follow a fixed pattern to address the dynamic data patterns (low-diversity or high-diversity patterns) often leading to performance degradation and consequently a decline in Quality of Experience (QoE). The primary issues we observed are:Different scenarios require different normalization methods (e.g., Instance Normalization is optimal in mixed domains but not in static domains). Model fine-tuning can potentially harm the model and waste time.Hence, it is crucial to design strategies for effectively measuring and managing distribution diversity to minimize its negative impact on model performance. Based on these observations, this paper proposes a new general method, named Diversity Adaptive Test-Time Adaptation (DATTA), aimed at improving QoE. DATTA dynamically selects the best batch normalization methods and fine-tuning strategies by leveraging the Diversity Score to differentiate between high and low diversity score batches. It features three key components: Diversity Discrimination (DD) to assess batch diversity, Diversity Adaptive Batch Normalization (DABN) to tailor normalization methods based on DD insights, and Diversity Adaptive Fine-Tuning (DAFT) to selectively fine-tune the model. Experimental results show that our method achieves up to a 21% increase in accuracy compared to state-of-the-art methodologies, indicating that our method maintains good model performance while demonstrating its robustness. Our code will be released soon.

Machine Learning

What problem does this paper attempt to address?

The paper aims to address the issue of handling distribution diversity in dynamic data streams using Test-Time Adaptation (TTA) methods. Specifically:

1. **Limitations of Traditional TTA Methods**: Existing TTA methods typically adopt a fixed pattern to cope with low-diversity or high-diversity data patterns, which leads to performance degradation, thereby affecting the Quality of Experience (QoE).
2. **Normalization Methods in Different Scenarios**: Different scenarios require different normalization methods (for example, Instance Normalization (IN) is optimal in mixed domains, but not in static domains).
3. **Risks of Model Fine-Tuning**: Model fine-tuning may damage the model and waste time.

To address these issues, the authors propose a new general method—Diversity Adaptive Test-Time Adaptation (DATTA), which aims to improve QoE by dynamically selecting the best batch normalization method and fine-tuning strategy. DATTA comprises three key components:
- **Diversity Discrimination (DD)**: Used to assess the diversity of batch data.
- **Diversity Adaptive Batch Normalization (DABN)**: Adjusts the normalization method based on the results of DD.
- **Diversity Adaptive Fine-Tuning (DAFT)**: Selectively performs model fine-tuning.

Experimental results show that the DATTA method improves accuracy by up to 21% on benchmark datasets compared to existing methods, demonstrating its robustness and effectiveness in handling dynamic data streams.