When Adversarial Training Meets Prompt Tuning: Adversarial Dual Prompt Tuning for Unsupervised Domain Adaptation

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised domain adaptation (UDA) seeks to transfer knowledge from a labeled source domain to an unlabeled target domain.
• Adversarial training is a common technique in UDA to minimize domain discrepancies.
• Prompt tuning offers an efficient approach for adapting large pre-trained vision-language models like CLIP.

Purpose of the Study:

• To propose a novel method, Adversarial Dual Prompt Tuning (ADAPT), for unsupervised domain adaptation.
• To leverage both text and visual prompts within a collaborative adversarial framework for CLIP model adaptation.
• To achieve domain alignment without relying on explicit domain discriminators.

Main Methods:

• ADAPT integrates text prompt tuning and visual prompt tuning into a collaborative adversarial framework.
• Text prompt tuning is designed to differentiate between source and target images.
• Visual prompt tuning aims to align the source and target domains by modeling joint probability distributions at global and category levels.

Main Results:

• Extensive experiments were conducted on four benchmark datasets.
• The proposed ADAPT method demonstrated significant effectiveness in unsupervised domain adaptation tasks.
• The approach successfully aligned domains by modeling joint probability distributions without explicit discriminators.

Conclusions:

• ADAPT offers an effective and novel approach to unsupervised domain adaptation using adversarial prompt tuning.
• The method successfully aligns domains by modeling joint probability distributions, eliminating the need for explicit domain discriminators.
• The collaborative adversarial game between text and visual prompts enhances model adaptability for CLIP.