Unsupervised Domain Adaptation With Class-Aware Memory Alignment.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised domain adaptation (UDA) aims to leverage labeled source data for unlabeled target domains.
• Current UDA methods often use mini-batch training, leading to unstable domain alignment due to random sampling.
• This instability can result in feature misalignment between domains.

Purpose of the Study:

• To propose a novel approach for robust unsupervised domain adaptation.
• To address the instability and misalignment issues in existing UDA methods.
• To enhance the transferability and discriminability of features during domain adaptation.

Main Methods:

• Introduced Class-Aware Memory Alignment (CMA), a novel UDA technique.
• Utilized two auxiliary class-aware memories to model source and target domain distributions.
• Implemented reliability-based filtering strategies to ensure memory quality.
• Developed a unified memory-based loss function for feature enhancement.

Main Results:

• CMA demonstrated superior performance compared to state-of-the-art (SOTA) methods.
• Ablation studies confirmed the effectiveness of the proposed CMA components.
• The memory-based approach achieved more stable and reliable domain alignment.

Conclusions:

• Class-Aware Memory Alignment (CMA) offers a robust solution for unsupervised domain adaptation.
• The proposed method effectively mitigates misalignment issues inherent in batch-level training.
• CMA enhances feature transferability and discriminability for improved cross-domain performance.