Source-Free Progressive Graph Learning for Open-Set Domain Adaptation.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Open-set domain adaptation (OSDA) faces challenges with irrelevant target classes and theoretical generalization bounds.
• Existing OSDA methods often require source and target data coexistence and struggle with accurate uncertainty estimation.
• The need for robust adaptation in source-free scenarios is critical for real-world applications.

Purpose of the Study:

• To propose the Progressive Graph Learning (PGL) framework to address limitations in current OSDA approaches.
• To develop a source-free OSDA (SF-OSDA) method that does not require simultaneous access to source and target domains.
• To improve the theoretical understanding of generalization bounds and prediction uncertainty in OSDA.

Main Methods:

• PGL decomposes the target hypothesis space and progressively pseudo-labels confident target samples using graph neural networks, episodic training, and adversarial learning.
• SF-PGL employs a two-stage framework with uniform selection of confident target instances and confidence-weighted classification loss for adaptation.
• The methods integrate graph neural networks with adversarial learning to bridge source and target domain distributions.

Main Results:

• PGL and SF-PGL demonstrate superior performance on image classification and action recognition benchmarks.
• The proposed methods show flexibility in recognizing both known and unknown target categories.
• Balanced pseudo-labeling significantly improves model calibration, reducing over- and under-confident predictions.

Conclusions:

• The PGL framework offers a robust solution for OSDA, including the challenging SF-OSDA setting.
• The approach provides theoretical guarantees on target error bounds and enhances prediction reliability.
• The findings advance the field of domain adaptation by improving accuracy, flexibility, and calibration.