TL-ADA: Transferable Loss-based Active Domain Adaptation.

Area of Science:

• Machine Learning
• Artificial Intelligence

Background:

• Active Domain Adaptation (ADA) aims to bridge the performance gap between supervised and unsupervised learning.
• Existing ADA research often focuses on query selection, neglecting effective training of new target samples.

Purpose of the Study:

• To propose a novel Transferable Loss-based Active Domain Adaptation (TL-ADA) framework.
• To address limitations of direct loss-based query selection in ADA, such as sample transferability and diversity issues.

Main Methods:

• Introduced a transferable doubly nested loss incorporating target pseudo-labels and domain adversarial loss.
• Developed a sequential training strategy considering domain type and label availability.
• Encouraged low self-entropy and diverse class distributions for pseudo-labels to enhance reliability.

Main Results:

• The proposed TL-ADA framework demonstrated superior performance compared to previous ADA methods on benchmark datasets.
• In-depth analysis validated the effectiveness of the novel transferable doubly nested loss and training strategy.

Conclusions:

• The TL-ADA framework offers an effective solution for training models in domain adaptation scenarios with limited labeled target data.
• The proposed methods improve sample selection and model training by addressing transferability and diversity challenges.