Progressive Hybrid Pseudo-Labeling for Unsupervised Domain Adaptation With Ascending Low-Rank Adaptation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised domain adaptation (UDA) using large vision-language models (VLMs) shows promise but struggles with noisy pseudo-labels and inefficient adaptation.
• Large domain shifts pose significant challenges to the generalization ability of current UDA methods.

Purpose of the Study:

• To propose a novel parameter-efficient paradigm, Progressive Hybrid Pseudo-Labeling with Ascending Low-Rank Adaptation (PHPL), for UDA.
• To address the challenges of noisy pseudo-labels and inefficient adaptation in large VLMs under significant domain shifts.

Main Methods:

• Introduced a progressive hybrid pseudo-labeling strategy with a teacher-student model and a weighting scheme to mitigate label noise.
• Developed an ascending low-rank adaptation (LoRA) strategy that allocates capacity depth-wise for efficient VLM adaptation.
• Implemented a progressive weighting scheme to transfer predictive responsibility from teacher to student, stabilizing self-training.

Main Results:

• PHPL achieved superior performance across five UDA benchmarks (Office-Home, Office-31, VisDA-2017, Mini-DomainNet, DomainNet).
• Demonstrated strong robustness and effectiveness on large-scale, challenging domain shift scenarios.
• Showcased significant reductions in computational overhead compared to existing solutions.

Conclusions:

• PHPL offers an effective and scalable lightweight adaptation paradigm for UDA with VLMs.
• The proposed methods successfully mitigate pseudo-label noise and enhance adaptation efficiency.
• PHPL represents a significant advancement in parameter-efficient UDA for large models.