# Visual Prompt Tuning for Test-time Domain Adaptation

**Authors:** Yunhe Gao, Xingjian Shi, Yi Zhu, Hao Wang, Zhiqiang Tang, Xiong Zhou, Mu Li, Dimitris N. Metaxas  
**arXiv:** [2210.04831](https://arxiv.org/abs/2210.04831)  
**Date:** 2022-10-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2210.04831v1/

## Key points

- Authors address the challenge of test-time adaptation (TTA) in machine learning models
- Emphasize the importance of models adapting to unseen data during testing
- Propose a method called data-efficient prompt tuning (DePT)
- DePT consists of two key components: incorporating visual prompts into the model and bootstrapping the source representation to the target domain using memory bank-based online pseudo labeling
- Hierarchical self-supervised regularization technique designed for prompts is optimized to mitigate error accumulation during self-training
- DePT achieves state-of-the-art performance on major adaptation benchmarks
- Shows promising results with as little as 1% or 10% of available data compared to using 100%
- Versatile and can be extended to online or multi-source TTA settings
- Offers improved accuracy and reduced resource consumption compared with existing methods.

## Summary

In this work titled "Visual Prompt Tuning for Test-time Domain Adaptation," authors Yunhe Gao, Xingjian Shi, Yi Zhu, Hao Wang, Zhiqiang Tang, Xiong Zhou, Mu Li and Dimitris N. Metaxas address the challenge of test-time adaptation (TTA) in machine learning models. They emphasize the importance of models being able to adapt to unseen data during testing to avoid performance degradation caused by distribution shifts in real-world deployment scenarios. To tackle this problem, the authors propose a method called data-efficient prompt tuning (DePT), which consists of two key components. First, DePT incorporates visual prompts into the vision Transformer model and only fine-tunes these prompts during adaptation to the target domain. This approach allows for efficient parameter tuning without overfitting to noise in the learning objective. The second component of DePT involves bootstrapping the source representation to the target domain using memory bank-based online pseudo labeling. Additionally, a hierarchical self-supervised regularization technique designed specifically for prompts is jointly optimized to mitigate error accumulation during self-training. The authors demonstrate that DePT achieves state-of-the-art performance on major adaptation benchmarks while also exhibiting superior data efficiency. Specifically, DePT shows promising results with as little as 1% or 10% of the available data compared to using 100% of the data for adaptation. Furthermore, DePT is versatile and can be extended to online or multi-source TTA settings. Overall, this work presents a practical and effective solution for test-time domain adaptation in machine learning models that offers improved accuracy and reduced resource consumption compared with existing methods.