Multi-Task Unified Domain Incremental Learning With Domain Difference Adapters.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Domain Incremental Learning (DIL) addresses adapting models to new data domains over time.
• Current DIL methods using prompts struggle with feature space adaptation due to limited representation.
• Catastrophic forgetting remains a challenge in continual learning scenarios.

Purpose of the Study:

• To propose a novel Domain Difference Adapters (DD-Adapters) method for effective domain adaptation in DIL.
• To enhance the feature space adaptation capabilities for multiple vision tasks.
• To mitigate catastrophic forgetting during the incremental learning process.

Main Methods:

• Proposing DD-Adapters that impose low-rank constraints on the base model to capture domain differences.
• Freezing the base model to preserve its cross-domain clustering ability.
• Introducing a prototype-guided domain selector (PDS) for dynamic adapter selection during inference.

Main Results:

• DD-Adapters demonstrate superior performance across object detection, instance segmentation, and image classification tasks.
• The method effectively adapts to new domain data distributions by leveraging low-rank properties of domain differences.
• Experimental evaluations on eight benchmark datasets confirm the method's effectiveness with minimal parameter overhead.

Conclusions:

• DD-Adapters offer a significant advancement in Domain Incremental Learning for vision tasks.
• The proposed approach effectively balances adaptation to new domains and knowledge retention.
• The method shows promise for real-world applications requiring continuous model adaptation.