Deep Label Propagation with Nuclear Norm Maximization for Visual Domain Adaptation.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Domain adaptation addresses distribution shifts between source and target datasets.
• Current methods often rely on pseudo-labeling for feature learning.
• Label Propagation (LP) is effective but underutilized in deep learning for domain adaptation and suffers from low confidence and class imbalance issues.

Purpose of the Study:

• To propose a novel domain adaptation approach, Deep Label Propagation with Nuclear Norm Maximization (DLP-NNM).
• To enhance label confidence and class diversity in LP for deep neural networks.
• To improve the reliability of predictions in the target domain for more effective feature learning.

Main Methods:

• Developed DLP-NNM incorporating nuclear norm maximization to improve LP.
• Designed an efficient algorithm to solve the optimization problem.
• Integrated the enhanced LP into a deep discriminative adaptation network using cross-entropy loss.

Main Results:

• The proposed DLP-NNM significantly enhances label confidence and class diversity.
• The method produces more reliable predictions for the target domain.
• Experimental results on three benchmark datasets show superior performance compared to state-of-the-art approaches.

Conclusions:

• DLP-NNM offers an effective solution for deep domain adaptation challenges.
• The approach successfully addresses limitations of traditional LP in deep learning contexts.
• This work advances the field of domain adaptation by improving feature learning and prediction accuracy.