Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Asymmetric Co-Training for Source-Free Few-Shot Domain Adaptation

Gengxu Li, Yuan Wu

TL;DR

This paper addresses domain adaptation under restrictive data access by introducing source-free few-shot domain adaptation (SFFSDA). It proposes asymmetric co-training (ACT), which fine-tunes a source model using a small $N$-way-$K$-shot labeled target support set and leverages a two-step optimization with label smoothing cross-entropy, entropy minimization, reverse cross-entropy, and a classifier determinacy disparity loss to align two target classifiers. ACT uses a two-branch architecture with strong data diversification via weak and AutoAugment based strong augmentations and optimizes with Sharpness-Aware Minimization, achieving superior performance over state-of-the-art SFUDA, transfer learning, and prior SFFSDA methods across multiple benchmarks. The results demonstrate that adapting a pre-trained model with limited labeled target data is practical and effective, handling long-tail and distribution shift scenarios while offering data-efficient, privacy-preserving domain adaptation with competitive accuracy. The work also provides ablations and analyses showing the contributions of each component and highlights potential directions for open-set and partial-set domain adaptation extensions.

Abstract

Source-free unsupervised domain adaptation (SFUDA) has gained significant attention as an alternative to traditional unsupervised domain adaptation (UDA), which relies on the constant availability of labeled source data. However, SFUDA approaches come with inherent limitations that are frequently overlooked. These challenges include performance degradation when the unlabeled target data fails to meet critical assumptions, such as having a closed-set label distribution identical to that of the source domain, or when sufficient unlabeled target data is unavailable-a common situation in real-world applications. To address these issues, we propose an asymmetric co-training (ACT) method specifically designed for the SFFSDA scenario. SFFSDA presents a more practical alternative to SFUDA, as gathering a few labeled target instances is more feasible than acquiring large volumes of unlabeled target data in many real-world contexts. Our ACT method begins by employing a weak-strong augmentation to enhance data diversity. Then we use a two-step optimization process to train the target model. In the first step, we optimize the label smoothing cross-entropy loss, the entropy of the class-conditional distribution, and the reverse-entropy loss to bolster the model's discriminative ability while mitigating overfitting. The second step focuses on reducing redundancy in the output space by minimizing classifier determinacy disparity. Extensive experiments across four benchmarks demonstrate the superiority of our ACT approach, which outperforms state-of-the-art SFUDA methods and transfer learning techniques. Our findings suggest that adapting a source pre-trained model using only a small amount of labeled target data offers a practical and dependable solution. The code is available at https://github.com/gengxuli/ACT.

Asymmetric Co-Training for Source-Free Few-Shot Domain Adaptation

TL;DR

This paper addresses domain adaptation under restrictive data access by introducing source-free few-shot domain adaptation (SFFSDA). It proposes asymmetric co-training (ACT), which fine-tunes a source model using a small -way--shot labeled target support set and leverages a two-step optimization with label smoothing cross-entropy, entropy minimization, reverse cross-entropy, and a classifier determinacy disparity loss to align two target classifiers. ACT uses a two-branch architecture with strong data diversification via weak and AutoAugment based strong augmentations and optimizes with Sharpness-Aware Minimization, achieving superior performance over state-of-the-art SFUDA, transfer learning, and prior SFFSDA methods across multiple benchmarks. The results demonstrate that adapting a pre-trained model with limited labeled target data is practical and effective, handling long-tail and distribution shift scenarios while offering data-efficient, privacy-preserving domain adaptation with competitive accuracy. The work also provides ablations and analyses showing the contributions of each component and highlights potential directions for open-set and partial-set domain adaptation extensions.

Abstract

Source-free unsupervised domain adaptation (SFUDA) has gained significant attention as an alternative to traditional unsupervised domain adaptation (UDA), which relies on the constant availability of labeled source data. However, SFUDA approaches come with inherent limitations that are frequently overlooked. These challenges include performance degradation when the unlabeled target data fails to meet critical assumptions, such as having a closed-set label distribution identical to that of the source domain, or when sufficient unlabeled target data is unavailable-a common situation in real-world applications. To address these issues, we propose an asymmetric co-training (ACT) method specifically designed for the SFFSDA scenario. SFFSDA presents a more practical alternative to SFUDA, as gathering a few labeled target instances is more feasible than acquiring large volumes of unlabeled target data in many real-world contexts. Our ACT method begins by employing a weak-strong augmentation to enhance data diversity. Then we use a two-step optimization process to train the target model. In the first step, we optimize the label smoothing cross-entropy loss, the entropy of the class-conditional distribution, and the reverse-entropy loss to bolster the model's discriminative ability while mitigating overfitting. The second step focuses on reducing redundancy in the output space by minimizing classifier determinacy disparity. Extensive experiments across four benchmarks demonstrate the superiority of our ACT approach, which outperforms state-of-the-art SFUDA methods and transfer learning techniques. Our findings suggest that adapting a source pre-trained model using only a small amount of labeled target data offers a practical and dependable solution. The code is available at https://github.com/gengxuli/ACT.

Paper Structure

This paper contains 29 sections, 8 equations, 5 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Unsupervised Domain Adaptation (UDA)
  4. Source-Free Unsupervised Domain Adaptation (SFUDA)
  5. Few-Shot Transfer Learning
  6. Method
  7. Overall Idea
  8. Data Diversity
  9. Model Structure
  10. Training Steps and Loss Functions
  11. Experiments
  12. Setup
  13. Datasets.
  14. Baseline Method
  15. Implementation Details
  16. ...and 14 more sections

Figures (5)

  • Figure 1: Comparison with SFUDA and transfer learning methods on OfficeHome dataset under 3-shot setting.
  • Figure 2: Overview of the ACT model structure and objective functions.
  • Figure 3: The label distributions of the three benchmarks: VisDA-C, Terra, and Office31. In each histogram, the x-axes and y-axes represent the class index and the proportion of this class in the domain, respectively
  • Figure 4: $t$-SNE visualization results on the Office-Home dataset (Ar $\rightarrow$ Cl(top) and Cl $\rightarrow$ Ar(bottom)). The left picture indicates the result from the source pre-trained model. The right picture indicates the result from the adapted model. Different colors represent different classes. The first $10$ classes are sampled for clear visualization.
  • Figure 5: $t$-SNE visualization results on the Office-Home dataset (Ar $\rightarrow$ Cl(top) and Cl $\rightarrow$ Ar(bottom)). The left picture indicates the result from the source pre-trained model. The right picture indicates the result from the adapted model. Different colors represent different classes. The first $10$ classes are sampled for clear visualization.