PaperlayerView exact revision on arXiv

Source-linked AI summary

Active Learning for Domain Adaptation: An Energy-Based Approach

Binhui Xie, Longhui Yuan, Shuang Li, Chi Harold Liu, Xinjing Cheng, Guoren Wang

arXiv:2112.01406v3cs.LGcs.CV

TL;DR

Domain adaptation still trails fully supervised learning, and efficient target labeling under domain shift remains difficult. EADA uses free-energy-based sampling and alignment to select informative target data, with theory and experiments supporting its effectiveness, including 2–6% gains over TQS across rounds on VisDA-2017 with ResNet-18.

  • Problem

    Domain adaptation often underperforms supervised learning, while conventional active learning is impractical for selecting informative labels under distribution shift.

  • Method

    EADA combines free-energy-based domain-characteristic sampling, MvSM instance-uncertainty sampling, and free-energy alignment for active domain adaptation.

  • Results

    EADA consistently outperforms alternative methods across labeling rounds, improving over TQS by 2–6% with ResNet-18 on VisDA-2017.

  • Takeaways & Limitations

    Energy-based sampling can select representative and informative target data while reducing the domain gap with limited labeling expenses.

Abstract

from arXiv · show

Unsupervised domain adaptation has recently emerged as an effective paradigm for generalizing deep neural networks to new target domains. However, there is still enormous potential to be tapped to reach the fully supervised performance. In this paper, we present a novel active learning strategy to assist knowledge transfer in the target domain, dubbed active domain adaptation. We start from an observation that energy-based models exhibit \textit{free energy biases} when training (source) and test (target) data come from different distributions. Inspired by this inherent mechanism, we empirically reveal that a simple yet efficient energy-based sampling strategy sheds light on selecting the most valuable target samples than existing approaches requiring particular architectures or computation of the distances. Our algorithm, Energy-based Active Domain Adaptation (EADA), queries groups of target data that incorporate both domain characteristic and instance uncertainty into every selection round. Meanwhile, by aligning the free energy of target data compact around the source domain via a regularization term, domain gap can be implicitly diminished. Through extensive experiments, we show that EADA surpasses state-of-the-art methods on well-known challenging benchmarks with substantial improvements, making it a useful option in the open world. Code is available at https://github.com/BIT-DA/EADA.

Introduction

Domain adaptation transfers knowledge from labeled source data to unlabeled target data, but its performance often remains below supervised learning. EADA addresses target labeling under domain shift by selecting samples using free-energy biases, domain characteristics, and instance uncertainty.

  • Introduction: Deep models require abundant labeled data, while domain adaptation transfers knowledge from labeled source to unlabeled target domains.These constraints arise in applications with massive unlabeled data or high labeling costs.
  • Introduction: Active domain adaptation seeks an efficient strategy for selecting informative target samples because conventional active learning is impractical under distribution shift.The central question is which target samples should be labeled to improve adaptation while limiting annotation cost.
  • Introduction: Energy-based models expose free-energy differences between source and target domains that can guide target-sample selection.Prior work cited in the paper also motivates energy models through calibration and improved in- versus out-of-distribution discrimination.
  • Introduction: EADA produces free-energy distributions better aligned than random selection and similar to full supervision on VisDA-2017.The full-supervision comparison labels all source and target data.
  • Introduction: EADA selects representative target data by combining free energy as a domain-characteristic signal with instance uncertainty.Higher free energy is treated as indicating greater dissimilarity from source data.
  • Introduction: The paper reports theoretical analysis, extensive experiments, and strong results with limited labeling expenses for energy-based active domain adaptation.Its stated contributions include a free-energy perspective, theoretical conditions, and empirical effectiveness.

Related Work

Prior active-learning methods emphasize uncertainty, representativeness, or hybrids, while active domain adaptation methods use learned domainness or clustering. EADA instead uses an efficient energy-based strategy that jointly captures domain characteristic and instance uncertainty without extra optimized parameters.

  • Active learning (AL): Conventional active learning commonly selects uncertain samples or representative sets through clustering and core-set selection.Hybrid methods combine these sampling objectives.
  • Active learning (AL): Conventional active-learning methods cannot directly handle the domain-shift issues of domain adaptation.EADA is presented as a response to this limitation.
  • Active domain adaptation: Active domain adaptation methods use domain discriminators or clustering, but domainness scores may be equally high for many target samples and fail to ensure distributional representativeness.The cited clustering approach selects active samples using target embeddings weighted by uncertainty.
  • Active domain adaptation: EADA uses energy to identify target samples unique to the target distribution and complementary to labeled source data.Its sampling protocol jointly accounts for domain characteristic and instance uncertainty.
  • Active domain adaptation: EADA requires no extra parameters to optimize and is described as efficient to learn.

Method

EADA trains an energy-based model with negative log-likelihood and free-energy alignment, then selects target samples using domain characteristic and instance uncertainty. Its theoretical analysis links positive gradient alignment to decreasing source free energy under supervised updates.

  • Problem setting: Active domain adaptation uses labeled source data and an unlabeled target pool, selecting a small active subset for annotation.The target pool is partitioned into labeled and unlabeled subsets during learning.
  • Energy-based Models Revisit: Energy-based models assign the lowest energy to the correct label and higher energy to incorrect labels.Classification inference can evaluate energies over the candidate label set and choose the minimum.
  • Energy-based Models Revisit: Free energy substitutes for unavailable target-sample probabilities when selecting low-probability, potentially representative target samples.The partition function is not reliably computable, motivating free energy as a practical surrogate.
  • Training process: EADA trains with negative log-likelihood and a free-energy alignment loss that reduces the higher free energy observed on target samples.The alignment loss is intended to align free-energy distributions across domains.
  • Selection process: Each selection round first chooses high-free-energy target candidates for domain characteristic, then samples uncertain points using MvSM.The overview specifies querying 1% of target samples per round and selecting MvSM points from an α1% candidate set.
  • Selection process: MvSM measures uncertainty as the energy difference between the lowest-energy and second-lowest-energy outputs, targeting instances near decision boundaries.The lowest and second-lowest outputs are denoted y∗ and y′, respectively.
  • Theoretical Analysis: For correctly classified source samples, a positive gradient inner product between negative log-likelihood and free energy implies that one supervised update decreases free energy.The deep-network theorem assumes smoothness, bounded gradients, and a specified learning-rate range.

Experiments

Across classification, segmentation, ablations, toy analysis, and efficiency tests, EADA consistently outperforms competing methods under limited target-label budgets. Results attribute this performance to combining domain-characteristic and uncertainty sampling with free-energy regularization.

  • Experimental setup: EADA is evaluated on image classification, semantic segmentation, toy domain-shift data, and multiple target-label budgets.Experiments cover VisDA-2017, Office-Home, Office-31, and GTAV→Cityscapes using pretrained ResNet-based models.
  • VisDA-2017: EADA surpasses baselines on VisDA-2017 with a 5% labeling budget and consistently outperforms alternatives across rounds.With ResNet-18, EADA improves over TQS by 2–6% across rounds; with ResNet-50, it is comparable to other methods at some rounds.
  • Office-Home and Office-31: EADA achieves the best performance across Office-Home and Office-31 tasks, with especially large margins on harder domain transfers.The reported benefit is attributed to jointly modeling domain characteristic and instance uncertainty while applying free-energy regularization.
  • GTAV→Cityscapes: On GTAV→Cityscapes, EADA substantially improves semantic-segmentation performance and becomes comparable to fully supervised training at the final round.The comparison uses 5% of target images queried for pixel-level labels of the whole image.
  • Ablation study: The full EADA method outperforms variants without free-energy sampling, uncertainty sampling, or free-energy regularization.Removing either sampling component reduces selection quality, while removing the regularizer weakens implicit domain-shift reduction.
  • Toy example: In the toy example, target error decreases from 52.0% for Source Only to 1.0% for EADA.Random and BADGE achieve 8.5% and 4.2% target error, respectively, while the example illustrates free-energy bias and redundant selection.
  • Selection-ratio analysis: EADA remains consistent across a wide range of selection ratios, although excluding either selection step causes performance to drop.The authors leave self-adaptive and weighted combinations of selection ratios for future work.
  • Efficiency: EADA obtains the best accuracy while being significantly more efficient than competitive baselines in query complexity and query time.BADGE and CLUE achieve strong mean accuracy but require a slower clustering step.

Conclusion

EADA addresses domain adaptation’s performance limitations with minimal labeling by selecting representative, informative target samples through energy-based sampling and regularization. Experiments across transfer scenarios support its effectiveness and suggest further value beyond fully supervised methods.

  • EADA selects limited target samples that are representative and informative using a novel energy-based sampling strategy.
  • A regularization term is used to implicitly diminish the domain gap.
  • The paper develops theoretical results describing when and why EADA is expected to work.
  • Experiments across various transfer scenarios demonstrate EADA’s effectiveness while targeting minimal label cost.

Appendix

The appendix includes dataset and implementation details, further analysis, additional results, and detailed theoretical proofs.

  • The appendix covers dataset details, implementation details, further analysis for EADA, additional results, and detailed theoretical proofs.

Dataset Details

VisDA-2017 is a large-scale synthetic-to-real image-classification dataset with over 280,000 images across 12 categories. This paper uses its synthetic training images as the source domain and real validation images as the target domain.

  • VisDA-2017 contains over 280,000 images from 12 categories for synthetic-to-real image classification.Its training, validation, and test sets contain 152,397 synthetic renderings, 55,388 real images, and 72,372 real images, respectively.
  • The paper uses VisDA-2017 training images as the source domain and validation images as the target domain.

Image classification.

Experiments use PyTorch with ImageNet-pretrained backbones matching competitive methods. Training uses AdaDelta for 50 epochs with batch size 32 and unified hyperparameters across three datasets.

  • Experiments use PyTorch and an ImageNet-pretrained backbone identical to those of competitive methods.This setup is intended to provide a fair comparison.
  • Training uses AdaDelta with learning rate 0.1 for 50 epochs and batch size 32.
  • The VisDA-2017, Office-Home, and Office-31 experiments share γ=0.01, α1=50, and α2=2.

Semantic segmentation.

The segmentation experiments use DeepLab-v2 with a ResNet-101 backbone and query 5% of target images for whole-image pixel-level labels.

  • Semantic segmentation.: DeepLab-v2 uses an ImageNet-pretrained ResNet-101 backbone and an Atrous Spatial Pyramid Pooling module.The ASPP sampling rates are fixed at {6, 12, 18, 24}.
  • Semantic segmentation.: The segmentation setup uses SGD with momentum 0.9 and weight decay 10^-4.

Labeling budget.

Each active domain adaptation round labels 1% of the target samples, with experiments using cumulative budgets of 5% or 20%.

  • Labeling budget.: Each selection round uses a labeling budget equal to 1% of all target samples.
  • Labeling budget.: Five rounds produce a total budget of 5% for VisDA-2017, Office-Home, and Office-31 final-model evaluations.
  • Labeling budget.: Twenty rounds produce a total budget of 20% on VisDA-2017, with accuracy reported after each round.

Baseline implementation.

The appendix describes baseline implementations, energy-based diagnostics, sampling behavior, regularization analysis, and loss-gradient interpretation for EADA.

  • Baseline implementation.: Random, BvSB, and entropy baselines select target samples using random choice, probability-margin uncertainty, or output entropy.
  • Baseline implementation.: AADA scores target samples with a domain-discriminator criterion combined with model entropy, while DBAL and TQS provide additional comparison methods.
  • Sampling behavior.: EADA selects samples from every class and obtains more representative and informative target instances across the first two VisDA-2017 rounds.
  • Free-energy analysis.: Free-energy distributions differ between source and target domains, whereas two source splits show no such bias; a regularization term is used to reduce these biases.
  • Regularization analysis.: EADA is robust across regularization weights because performance first increases and then decreases slightly over the tested range.The tested weights are γ ∈ {0.001, 0.005, 0.01, 0.05, 0.1}.
  • Gradient analysis.: The negative log-likelihood gradient pulls down the correct-answer energy while the second term pulls down incorrect-answer energies in the analyzed formulation.
Loading 2112.01406v3…