Local Uncertainty Energy Transfer for Active Domain Adaptation

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Active Domain Adaptation (ADA) aims to improve knowledge transfer from labeled source to unlabeled target domains by selecting informative target samples.
• Existing ADA methods often overlook local sample uncertainty, leading to the selection of detrimental anomalous samples.
• This limitation hinders model performance and efficient knowledge transfer.

Purpose of the Study:

• To introduce a novel Active Domain Adaptation approach, Local Uncertainty Energy Transfer (LUET).
• To address the issue of anomalous sample selection in target domains by considering local uncertainty.
• To improve the efficiency and accuracy of knowledge transfer in domain adaptation.

Main Methods:

• LUET integrates active learning with local uncertainty and energy transfer alignment constraints.
• The active learning module employs local uncertainty energy and entropy-weighted class confusion for sample selection, avoiding anomalies.
• A global and local energy-transfer alignment module is utilized to bridge the domain gap, complemented by negative log-likelihood loss.

Main Results:

• LUET effectively selects informative and representative target samples by considering local uncertainty.
• The proposed method successfully mitigates the negative impact of anomalous samples on model training.
• Experiments demonstrate LUET's superior performance over state-of-the-art approaches on multiple domain-adaptive datasets.

Conclusions:

• LUET offers a robust framework for Active Domain Adaptation by incorporating local uncertainty and energy transfer.
• The method enhances knowledge transfer efficiency and model accuracy by intelligently selecting target samples.
• LUET represents a significant advancement in addressing domain shift challenges in machine learning.