Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Energy-Based Pseudo-Label Refining for Source-free Domain Adaptation

Xinru Meng, Han Sun, Jiamei Liu, Ningzhong Liu, Huiyu Zhou

TL;DR

This work tackles the challenge of source-free domain adaptation by reducing pseudo-label noise through energy-based refinement. It introduces EBPR, a framework that generates pseudo-labels via transport-cost-based clustering, applies adaptive energy thresholds to filter labels, and uses enhanced consistency training for hard samples, all without access to source data. Empirical results on Office-31, Office-Home, and VisDA-C demonstrate state-of-the-art performance and improved handling of class-imbalance and noisy labels. The approach offers a practical, robust pathway for SFDA, though it incurs higher computational cost on very large datasets.

Abstract

Source-free domain adaptation (SFDA), which involves adapting models without access to source data, is both demanding and challenging. Existing SFDA techniques typically rely on pseudo-labels generated from confidence levels, leading to negative transfer due to significant noise. To tackle this problem, Energy-Based Pseudo-Label Refining (EBPR) is proposed for SFDA. Pseudo-labels are created for all sample clusters according to their energy scores. Global and class energy thresholds are computed to selectively filter pseudo-labels. Furthermore, a contrastive learning strategy is introduced to filter difficult samples, aligning them with their augmented versions to learn more discriminative features. Our method is validated on the Office-31, Office-Home, and VisDA-C datasets, consistently finding that our model outperformed state-of-the-art methods.

Energy-Based Pseudo-Label Refining for Source-free Domain Adaptation

TL;DR

This work tackles the challenge of source-free domain adaptation by reducing pseudo-label noise through energy-based refinement. It introduces EBPR, a framework that generates pseudo-labels via transport-cost-based clustering, applies adaptive energy thresholds to filter labels, and uses enhanced consistency training for hard samples, all without access to source data. Empirical results on Office-31, Office-Home, and VisDA-C demonstrate state-of-the-art performance and improved handling of class-imbalance and noisy labels. The approach offers a practical, robust pathway for SFDA, though it incurs higher computational cost on very large datasets.

Abstract

Source-free domain adaptation (SFDA), which involves adapting models without access to source data, is both demanding and challenging. Existing SFDA techniques typically rely on pseudo-labels generated from confidence levels, leading to negative transfer due to significant noise. To tackle this problem, Energy-Based Pseudo-Label Refining (EBPR) is proposed for SFDA. Pseudo-labels are created for all sample clusters according to their energy scores. Global and class energy thresholds are computed to selectively filter pseudo-labels. Furthermore, a contrastive learning strategy is introduced to filter difficult samples, aligning them with their augmented versions to learn more discriminative features. Our method is validated on the Office-31, Office-Home, and VisDA-C datasets, consistently finding that our model outperformed state-of-the-art methods.

Paper Structure

This paper contains 20 sections, 11 equations, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Pseudo-labels methods
  4. Energy-based models
  5. Proposed Methodology
  6. Problem definition
  7. Method
  8. Overview of the training process
  9. Pseudo-label generation with transport cost
  10. Adaptive energy threshold
  11. Enhanced consistency training
  12. Experimental Evaluations
  13. Datasets
  14. Model architecture and experimental protocols
  15. Comparison to the literature and baselines
  16. ...and 5 more sections

Figures (3)

  • Figure 1: The framework of the proposed EBPR.
  • Figure 2: Influence of $\rho$ in C → A accuracy (%) on the Office-Home dataset.
  • Figure 3: The t-SNE of the task A → W. The first row shows the visualization of category features where the different colors denote different categories. The second row shows features of the filtered intermediate samples (in red) and others (in blue).