PaperlayerView exact revision on arXiv

Source-linked AI summary

Active Domain Adaptation via Clustering Uncertainty-weighted Embeddings

Viraj Prabhu, Arjun Chandrasekaran, Kate Saenko, Judy Hoffman

arXiv:2010.08666v3cs.CVcs.LG

TL;DR

Active Domain Adaptation asks how to select a cost-effective subset of target labels for adapting a source model under domain shift. CLUE performs uncertainty-weighted clustering of target embeddings to choose instances that are both informative and diverse. Across six diverse image-classification shifts and multiple learning settings, CLUE consistently outperforms competing acquisition strategies.

  • Problem

    Active Domain Adaptation needs cost-effective selection of target labels because uncertainty-only and diversity-only active-learning strategies are less effective under domain shift.

  • Method

    CLUE weights target embeddings by predictive uncertainty, clusters them in feature space, and selects representative instances near cluster centroids for labeling.

  • Results

    CLUE consistently performs best across learning strategies, shifts, benchmarks, and most rounds, improving over prior Active DA by as much as 9% in some cases.

  • Takeaways & Limitations

    Combining uncertainty and diversity provides a versatile label-acquisition strategy for Active DA across diverse domain shifts.

  • Takeaways & Limitations

    The method assumes source instances have low predictive entropy, an assumption generally associated with supervised training.

Abstract

from arXiv · show

Generalizing deep neural networks to new target domains is critical to their real-world utility. In practice, it may be feasible to get some target data labeled, but to be cost-effective it is desirable to select a maximally-informative subset via active learning (AL). We study the problem of AL under a domain shift, called Active Domain Adaptation (Active DA). We demonstrate how existing AL approaches based solely on model uncertainty or diversity sampling are less effective for Active DA. We propose Clustering Uncertainty-weighted Embeddings (CLUE), a novel label acquisition strategy for Active DA that performs uncertainty-weighted clustering to identify target instances for labeling that are both uncertain under the model and diverse in feature space. CLUE consistently outperforms competing label acquisition strategies for Active DA and AL across learning settings on 6 diverse domain shifts for image classification.

1. Introduction

Active Domain Adaptation selects a limited budget of target labels to adapt a source model under domain shift. CLUE combines uncertainty-weighted feature-space clustering to select informative, diverse, and representative instances, outperforming competing strategies across six shifts.

  • Problem: Active DA selects target instances for labeling from an unlabeled target domain to adapt a source-trained model under a fixed budget.The task targets high accuracy on the target test set when target labels are costly or limited.
  • Limitations of existing methods: Uncertainty-only and diversity-only sampling are less effective for Active DA because domain shift can produce redundant, uninformative, or already-aligned selections.Uncertainty may be miscalibrated, while diversity can sample regions that are already well aligned across domains.
  • CLUE: CLUE weights target embeddings by model uncertainty and clusters them to select instances that are both uncertain and diverse in feature space.The method combines informative sampling with reduced redundancy without complex gradient or domain-discriminator diversity measures.
  • Results: CLUE improves on prior Active DA and active-learning methods across six diverse image-classification domain shifts and multiple learning strategies.The reported improvement over prior Active DA reaches as much as 9% in some cases.

2. Related Work

Prior work studies uncertainty sampling, diversity sampling, domain adaptation, and their combinations, but Active DA requires selecting target labels specifically for adaptation under domain shift. CLUE combines uncertainty-weighted embeddings with clustering for this setting.

  • Active Learning: Uncertainty sampling selects instances with high model uncertainty, while diversity sampling selects representative and dissimilar instances in a learned embedding space.Common uncertainty measures include entropy and classification margins; diversity methods often use clustering or feature-space coverage.
  • Hybrid methods: Hybrid active-learning methods combine uncertainty and diversity through bandit selection, uncertainty pre-filtering, or clustering gradient embeddings.BADGE, for example, clusters hallucinated gradient embeddings.
  • Domain Adaptation: Domain adaptation transfers models from labeled source domains to unlabeled or partially labeled target domains using discrepancy minimization or adversarial alignment.These methods typically focus on aligning source and target feature spaces rather than selecting target labels.
  • Active Domain Adaptation: Active DA differs from semi-supervised domain adaptation by actively selecting which target instances receive labels instead of using a random labeled subset.Earlier Active DA work combined uncertainty with domain separation or optimized source importance weights.
  • CLUE: CLUE applies uncertainty-weighted clustering to target embeddings, selecting diverse instances for labeling in the Active DA setting.Its pipeline weights embeddings by uncertainty, clusters them, and chooses representative points near cluster centroids.

3. Approach

CLUE jointly models informativeness and diversity by weighting target feature embeddings with predictive uncertainty and clustering them. It then acquires representative points near cluster centroids and updates the model through active domain adaptation.

  • Problem setup: Active DA queries an oracle for a budgeted subset of unlabeled target instances and uses the acquired labels to improve target-domain performance.The setting includes labeled source data, unlabeled target data, and a per-round labeling budget.
  • Uncertainty weighting: CLUE uses predictive entropy as an informativeness measure and weights feature-space clustering so uncertain regions receive greater emphasis.Entropy is defined from the model’s predictive class distribution and can also reflect target-domain uncertainty under domain shift.
  • Joint objective: The uncertainty-weighted clustering objective jointly captures feature-space diversity and model uncertainty instead of optimizing diversity alone.The weighted objective emphasizes representative sampling from uncertain regions while preserving low within-cluster variance.
  • Optimization and selection: CLUE approximates its NP-hard objective with weighted K-Means, setting the number of clusters equal to the labeling budget.It uses penultimate-layer CNN activations and selects the nearest instance to each weighted cluster mean to avoid outliers.
  • Model update: The selected target labels, optionally combined with labeled source and unlabeled target data, update the model through finetuning or domain-adaptation strategies.The full procedure iterates acquisition and semi-supervised domain alignment across rounds.

4. Experiments

The experiments evaluate CLUE against uncertainty-, diversity-, and hybrid sampling methods across six image-classification domain shifts, learning strategies, and budgets. CLUE generally performs best, while visualizations and ablations show that uncertainty-weighted clustering selects diverse points from dense, uncertain regions.

  • Experimental setup: The evaluation covers six domain shifts across DomainNet, DIGITS, and Office, using finetuning, MME, and DANN-based adaptation settings.DomainNet includes four shifts of increasing difficulty; DIGITS uses SVHN→MNIST, and Office uses DSLR→Amazon.
  • Experimental setup: The study compares CLUE with AADA and AL baselines based on uncertainty, diversity, or hybrid sampling under standard batch active learning.The baselines include entropy, margin, coreset, and BADGE, with experiments reporting target-test accuracy across intermediate budgets.
  • Main results: Uncertainty- or diversity-only methods frequently underperform random sampling on challenging shifts, whereas hybrid methods such as CLUE and BADGE are more versatile.Uncertainty methods work well on some easier shifts, but generalize poorly overall to difficult shifts such as S→P and C→Q.
  • Main results: CLUE consistently performs best across learning strategies, shifts, benchmarks, and most rounds, outperforming margin and coreset at B = 2k on four DomainNet shifts.When finetuning, the gains are 1.4% over margin and 3% over coreset; with MME, they are 1.3% and 2.3%, respectively.
  • Analysis and ablations: The SVHN→MNIST visualization shows entropy selecting redundant points, coreset missing already aligned regions, and CLUE selecting dissimilar points from dense, high-uncertainty regions.Ablation results further show that entropy-based uncertainty weighting outperforms uniform weighting and narrowly beats margin weighting on C→S.
  • Main results: At B = 30 on SVHN→MNIST, CLUE improves over margin, coreset, and BADGE by 1.9%, 12.3%, and 5.2% with finetuning.On DSLR→Amazon, the corresponding improvements are 4%, 2.5%, and 0.6%.
  • Learning strategies: MME consistently outperforms finetuning across active-learning methods, while CLUE combined with DANN outperforms AADA by 0.4%–2% on DomainNet.The performance gap over AADA increases with shift difficulty, reaching a 3.4% gain at B = 2k on C→Q.

5. Conclusion

The paper addresses Active Domain Adaptation by selecting target instances that are both uncertain and diverse through uncertainty-weighted clustering. CLUE outperforms competing strategies across learning settings and domain shifts.

  • CLUE selects diverse, informative target instances from dense feature-space regions by clustering embeddings weighted by model uncertainty.This combines uncertainty and diversity for label acquisition under domain shift.

6. Appendix

The appendix analyzes CLUE across traditional active learning, longer acquisition schedules, and comparisons with BADGE. CLUE is strongest early, generally outperforms competing methods, and narrows its advantage as labeling increases.

  • Traditional active learning: CLUE significantly outperforms prior active learning methods on DomainNet C→S with ImageNet initialization and performs on-par with state-of-the-art methods on SVHN from scratch.On SVHN, CLUE is also significantly better than uniform sampling over most rounds.
  • When gains saturate: Performance differences across methods narrow after roughly 15k labels, following CLUE’s strongest improvements during the initial training stages.The saturation check uses 40 rounds with 500 labels per round on Clipart→Sketch.
  • Comparing CLUE and BADGE: CLUE and BADGE differ in feature space, uncertainty measure, and clustering procedure.CLUE uses uncertainty-scaled penultimate-layer embeddings and uncertainty-weighted K-Means, whereas BADGE uses gradient embeddings and KMeans++.
  • Comparing CLUE and BADGE: BADGE’s gradient embeddings reach approximately 176k dimensions on DomainNet C→S, while CLUE uses 512-dimensional embeddings.The appendix connects the lower-dimensional CLUE representation to more meaningful distance-based clustering.
  • Comparing CLUE and BADGE: Across 40 DomainNet settings with MME, CLUE performs as well or better than BADGE in 38 settings and better in 24.On other benchmarks, CLUE gains can be larger, although BADGE and CLUE can perform similarly in later rounds on some shifts.

6.4. Dataset details

The experiments use DomainNet as the primary multi-domain benchmark and also evaluate active learning on SVHN. The appendix provides dataset scale, domain coverage, and representative qualitative or statistical references.

  • DomainNet: DomainNet contains 0.6 million images across six domains, with experiments using four shifts from Real, Clipart, Sketch, Painting, and Quickdraw.Real is not used as a target domain because models use ImageNet initialization.
  • Active learning benchmarks: Figure 6 reports active learning performance for CLUE on DomainNet C→S with ImageNet initialization and SVHN with finetuning from scratch.The figure caption states that CLUE significantly outperforms state-of-the-art methods in the first setting and performs on-par in the second.
  • Dataset statistics: Table 5 summarizes DomainNet train and test statistics for the evaluated domains.
  • DIGITS: The SVHN benchmark uses grayscale digit images, with representative examples shown in Figure 7.The supplied dataset description states that SVHN contains 70k images: 60k training and 10k test.

6.5. Code and Implementation Details

The implementation uses weighted K-Means for CLUE and tunes an optional softmax temperature on a small validation set. Baseline implementations include BADGE’s gradient embeddings and AADA’s discriminator-based acquisition score.

  • CLUE implementation: CLUE uses weighted K-Means with K-means++ initialization and the Elkan algorithm, whose approximate time complexity is O(nke).The implementation uses 300 K-Means iterations in the experiments.
  • Temperature tuning: The softmax temperature is tuned to T=0.1 on 482 Clipart→Sketch examples and then used across DomainNet shifts.The appendix describes temperature tuning as optional when a small validation set is available and reports strong results with the default T=1.0.
  • DIGITS implementation: DIGITS experiments use a modified LeNet architecture with Adam, learning rate 2x10^-4, weight decay 10^-5, batch size 128, and 60 epochs per round.The learning rate is halved every 20 epochs.
  • Baseline implementations: BADGE forms gradient embeddings from classifier-weight gradients and applies K-Means++, while AADA scores target points using domain-discriminator targetness and model entropy.AADA selects B instances randomly from the top 2% of scores to reduce redundancy and adds entropy minimization.

6.6. CLUE: Qualitative Analysis

The qualitative analysis compares how entropy, coreset, and CLUE select target instances, and visualizes CLUE’s sampling behavior across rounds. CLUE combines concentration on difficult examples with coverage across classes and feature-space regions.

  • MME aligns some classes without target labels, but substantial misalignments remain for classes 0, 4, and 6.
  • With B = 30, entropy selects many class-8 examples that appear challenging to humans, while coreset spreads selections more broadly across classes.
  • CLUE combines entropy’s emphasis on class 8 with coreset’s broader coverage by selecting at least a few instances from every class.
  • Across rounds, t-SNE visualizations show CLUE sampling target feature-space regions using correct and incorrect model predictions as reference points.

6.7. Extended Description of the CLUE Objective

The extended objective formulates representative target-set selection as variance-minimizing clustering in CNN feature space. CLUE then weights instances by model uncertainty so uncertain examples exert greater influence on cluster representatives.

  • CLUE measures dissimilarity within each target subset using variance in the CNN representation space.A small σ2(X_k) indicates that instances in X_k are similar to one another.
  • The unweighted objective partitions unlabeled target data into K sets and minimizes the sum of their within-set variances.
  • Weighted variance gives more informative or uncertain instances greater influence when identifying representative instances.
  • CLUE uses target-model uncertainty H(Y|x) as each instance’s informativeness weight in the weighted-variance partitioning objective.The weighted cluster center is computed from uncertainty-weighted feature embeddings.
  • The extended derivation concludes with CLUE’s overall set-partitioning objective based on uncertainty-weighted feature-space variance.

6.8. Full performance plots

The appendix provides full Active DA learning curves corresponding to the main-paper tables. Results cover three learning strategies and report means with one standard deviation across three runs.

  • The full plots complement intermediate-budget results reported in Tables 1 and 2 of the main paper.
  • The experiments compare finetuning, semi-supervised DA via MME, and semi-supervised DA via DANN, each starting from a source model.
  • Learning curves report performance means and one standard deviation over three experimental runs, with standard deviation shown by shading.

6.9. Future Work

The authors identify future work in uncertainty modeling, few-shot architectures, and extending Active DA beyond image classification. The appendix figures provide full plots across several domain-shift settings and learning strategies.

  • Future versions of CLUE could replace model entropy with alternative uncertainty measures, including deep-ensemble uncertainty shown to be more reliable under dataset shift.
  • Specialized few-shot-learning architectures could be incorporated to address label sparsity in the target domain.
  • Active domain adaptation remains to be studied for object detection and semantic segmentation, beyond this paper’s image-classification setting.
  • The DomainNet appendix plots cover four shifts across five domains and compare CLUE with uncertainty-, diversity-, and Active DA baselines.
  • The DIGITS and Office appendix plots compare finetuning, MME, and DANN learning strategies, reporting means and one standard deviation over three runs.
Loading 2010.08666v3…