PaperlayerView exact revision on arXiv

Source-linked AI summary

Spawrious: A Benchmark for Fine Control of Spurious Correlation Biases

Aengus Lynch, Gbètondji J-S Dovonon, Jean Kaddour, Ricardo Silva

arXiv:2303.05470v3cs.CVcs.LG

TL;DR

Existing benchmarks focus on O2O correlations and do not test reversed M2M correlations, while reliance on spurious features is a concern in out-of-distribution settings. The paper builds Spawrious, an image benchmark with O2O and M2M correlations across three difficulty levels, using advances in text-to-image and image-captioning models. On Spawrious Hard-splits, state-of-the-art methods achieve <70% accuracy using ResNet50 pretrained on ImageNet.

  • Problem

    Existing benchmarks focus on O2O correlations and do not test reversed M2M correlations, while reliance on spurious features is a concern in out-of-distribution settings.

    Source: §1 · pp. 1–2

  • Method

    The paper builds Spawrious, an image benchmark with O2O and M2M correlations across three difficulty levels, using advances in text-to-image and image-captioning models.

    Source: §8 · p. 9

  • Results

    On Spawrious Hard-splits, state-of-the-art methods achieve <70% accuracy using ResNet50 pretrained on ImageNet.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Model-error analysis revealed dependence on irrelevant backgrounds, underscoring the dataset's difficulty and the need for further investigation.

    Source: §8 · p. 9

    The benchmark's stated future directions include instantiating desiderata for non-background spurious attributes and other data modalities.

    Source: §7 · p. 9

Abstract

from arXiv · show

The problem of spurious correlations (SCs) arises when a classifier relies on non-predictive features that happen to be correlated with the labels in the training data. For example, a classifier may misclassify dog breeds based on the background of dog images. This happens when the backgrounds are correlated with other breeds in the training data, leading to misclassifications during test time. Previous SC benchmark datasets suffer from varying issues, e.g., over-saturation or only containing one-to-one (O2O) SCs, but no many-to-many (M2M) SCs arising between groups of spurious attributes and classes. In this paper, we present \benchmark-\{O2O, M2M\}-\{Easy, Medium, Hard\}, an image classification benchmark suite containing spurious correlations between classes and backgrounds. To create this dataset, we employ a text-to-image model to generate photo-realistic images and an image captioning model to filter out unsuitable ones. The resulting dataset is of high quality and contains approximately 152k images. Our experimental results demonstrate that state-of-the-art group robustness methods struggle with \benchmark, most notably on the Hard-splits with none of them getting over $70\%$ accuracy on the hardest split using a ResNet50 pretrained on ImageNet. By examining model misclassifications, we detect reliances on spurious backgrounds, demonstrating that our dataset provides a significant challenge.

1 Introduction

Spawrious targets gaps in spurious-correlation benchmarks, especially their limited treatment of many-to-many (M2M) correlations. It introduces O2O and M2M image-classification challenges at three difficulty levels, and reports that state-of-the-art methods struggle most on the Hard splits.

1 Introduction

  • The paper identifies reliance on spurious, non-causal features as a concern for models evaluated on out-of-distribution settings.It cites background reliance in ImageNet classifiers as an example.Source: §1 · p. 1
  • Existing benchmarks focus on O2O correlations, whereas M2M correlations can reverse across groups of classes and backgrounds in ways that cannot be matched class by class.The paper presents M2M correlations as a qualitatively different structure, not an aggregation of O2O correlations.Source: §1 · pp. 1–2
  • Spawrious introduces O2O and M2M image-classification datasets, each with Easy, Medium, and Hard difficulty levels.The suite contains 152,064 images at 224 × 224 resolution and is designed to probe classifier reliance on spurious correlations.Source: §1 · p. 2
  • State-of-the-art methods achieve <70% accuracy on Spawrious Hard splits using an ImageNet-pretrained ResNet50.The authors also report that probing misclassifications provides evidence of reliance on spurious features, and that gains from larger architectures are inconsistent across methods.Source: §1 · p. 2

2 Existing Benchmarks

The paper contrasts Spawrious with prior benchmarks that either do not explicitly reverse spurious correlations at test time or lack M2M environments. It also reports a larger gap between robustness methods on Spawrious than the near-parity observed in DomainBed.

2 Existing Benchmarks

  • The paper describes DomainBed as a domain-generalization suite, not a benchmark focused on spurious correlations, and reports that its methods perform about the same on average.The authors raise whether some DomainBed challenges, such as photo-realistic training to cartoon test shifts, are well-posed.Source: §2 · p. 2
  • ERM is 7.30 percentage points worse than the best method on average on Spawrious, where methods show stark differences.The authors contrast this with DomainBed, where ERM and domain-generalization methods perform about the same on average after hyperparameter tuning.Source: §2 · pp. 2–3
  • Table 1 compares Spawrious and other benchmarks by whether they pose domain-generalization, O2O, and M2M challenges.The caption states the comparison dimensions but does not provide the table's cell values.Source: Table 1 · p. 3
  • None of the prior benchmarks discussed includes explicit training and test environments for M2M spurious correlations.The paper distinguishes these from benchmarks that collect data across environments without inducing explicit spurious correlations between environments and labels, and from those lacking unseen test domains.Source: §2 · p. 3

3 Benchmark Desiderata

The section sets out six desiderata for a benchmark and formalizes O2O and M2M spurious correlations as train–test distribution shifts. Its M2M formulation captures reversed associations between groups of attributes and classes.

3 Benchmark Desiderata

  • The section motivates the benchmark by seeking to address shortcomings in existing benchmarks and formalize the correlation types under study.Source: §3 · p. 3

3.1 Six Desiderata

  • The benchmark desiderata are photo-realism, non-binary classification, inter-class homogeneity with intra-class heterogeneity, high-fidelity backgrounds, multiple training environments, and multiple difficulty levels.These criteria aim to support non-trivial classification under realistic conditions and allow study of cost trade-offs across difficulty.Source: §3.1 · pp. 3–4

3.2 Spurious Correlations (One-To-One)

  • A spurious feature is harmful in the described challenge when it alone can trigger a class prediction without the class object present.The section defines correlated non-causal features as features that frequently occur with a class without causing its appearance; they are spurious when the classifier perceives them as a feature of that class.Source: §3.2 · p. 4
  • O2O spurious correlations pair spurious features with labels in training environments, with the correlations differing in the test environment.The paper frames evaluation around performance when the training correlations are reverted.Source: §3.2 · p. 4

3.3 Many-To-Many Spurious Correlations

  • M2M spurious correlations link disjoint groups of attributes and classes in training, then reverse those group-level correlations at test time.In the example, each class in a group appears with each background in its corresponding group in equal proportion during training.Source: §3.3 · p. 4
  • The paper presents M2M correlations as richer structures that cannot be expressed or decomposed as O2O correlations.Figure 2 illustrates the distributions through class-group and spurious-attribute-group hierarchies.Source: §3.3 · p. 4

4 The Spawrious Challenge

Spawrious instantiates synthetic dog-breed/background classification challenges with O2O and M2M spurious correlations. It combines text-to-image generation and caption filtering, then varies class-background combinations across difficulty splits.

4 The Spawrious Challenge

  • Spawrious comprises synthetic images of four dog breeds across six background locations, totaling 152,064 images.The benchmark covers O2O and M2M spurious correlations at Easy, Medium, and Hard difficulty levels.Source: §4 · p. 5, §4.2 · p. 6

4.1 Dataset Construction

  • The construction pipeline generates images from sampled prompts with Stable Diffusion v1.4 and retains images whose captions contain important class keywords.A ViT-GPT2 captioning model filters out images without detectable class objects.Source: §4.1 · p. 5

4.2 Satisfying Desiderata

  • In O2O splits, each class has a spurious background seen only with that class and a generic background shared equally across classes.Two training environments use distinct proportions of spurious and generic backgrounds.Source: §4.2 · p. 6
  • M2M splits form disjoint groups of two classes and two backgrounds, pairing each class with backgrounds from its group.The groups are swapped across two environments.Source: §4.2 · p. 6
  • Easy, Medium, and Hard differ in available class-background combinations; final splits were selected using empirical ERM performance, without a theoretical difficulty rationale.The remaining combinations were selected through 20 random trials each for O2O and M2M.Source: §4.2 · p. 6

5 Experiments

The experiments evaluate group-robustness methods on test environments where training spurious correlations change. Most methods struggle more as challenges become harder, and ERM errors align with spurious backgrounds.

5 Experiments

  • The evaluation uses six popular group-robustness methods and tests changed-correlation environments; O2O uses one randomly selected test combination, while M2M swaps the two class and background groups.Source: §5 · pp. 6–7
  • The main results use ImageNet-pretrained ResNet18 models; hyperparameters are tuned through random search, with selected settings reused across related splits.Source: §5.1 · p. 7, §5 · p. 7

5.1 Results

  • With ResNet18, the top three listed methods perform significantly worse than the bottom three; with ResNet50, the hierarchy remains except MMD-AAE worsens.Source: §5.1 · p. 7
  • 11.36%: IRM performs worst on average, below the best method MMD-AAE; most methods worsen as challenges become harder, especially M2M Medium and Hard.Source: §5.1 · p. 7

5.2 Misclassification analysis

  • ERM misclassifies Bulldog test images as classes associated with the same backgrounds in training, including Mountains and Snow cases shown for the two Hard challenges.Source: §5.2 · p. 8, Figure 4 · p. 8
  • Class counts are balanced across classes and groups in training and test data, yet the authors conjecture that ERM relies heavily on Mountains and Snow backgrounds.Source: §5.2 · p. 8
  • The confusion matrix includes Dachshund images misclassified as Labrador; the authors conjecture this reflects Snow being Labrador’s most common training background.Source: §5.2 · p. 8, Figure 5 · p. 8

6 Related Work

Related work distinguishes distribution-shift failures caused by group imbalance from the balanced-class spurious correlations studied here, and describes causal perspectives behind invariance learning.

6 Related Work

  • Geometric skew is group imbalance across data types that induces spurious correlations and can cause misclassification when group balance changes.Source: §6 · p. 8
  • The paper studies spurious correlations that persist despite perfect class balance among groups, contrasting them with imbalance-based distribution-shift failures.Source: §6 · p. 8
  • Causal perspectives formalize environment-specific data as interventional distributions and motivate invariance-learning methods that make assumptions about the data-generating process.Source: §6 · p. 9

7 Limitations and Future Work

The authors identify non-background spurious attributes as a direction for instantiating their desiderata. They also propose extending the benchmark to other modalities and evaluating additional generalization techniques.

7 Limitations and Future Work

  • Instantiating the desiderata with non-background spurious attributes is identified as future work.The example given is a spurious feature that is part of the classification object rather than its background.Source: §7 · p. 9
  • Text classification using text-generation capabilities of large language models is proposed as another modality for future instantiation.Source: §7 · p. 9
  • The authors propose evaluating additional generalization approaches, including robustness penalties, meta-learning, domain adaptation, and data augmentation.Source: §7 · p. 9

8 Conclusion

The conclusion presents Spawrious as an image-classification benchmark covering O2O and M2M spurious correlations. It reports that group-robustness methods struggle, especially on hard splits, and that model errors reveal reliance on irrelevant backgrounds.

8 Conclusion

  • Spawrious is an image-classification benchmark with one-to-one and many-to-many spurious correlations.Source: §8 · p. 9
  • The benchmark comprises six datasets designed using text-to-image and image-captioning models.Source: §8 · p. 9
  • Below 70% accuracy on hard splits, state-of-the-art group-robustness techniques remain insufficient on Spawrious.Source: §8 · p. 9
  • Analysis of model errors revealed dependence on irrelevant backgrounds, and the authors call for further investigation.Source: §8 · p. 9

A Effect of Model Architectures

The architecture experiments report that the best CORAL ResNet50 result exceeds the best ViT by 17.1%, while the ViT-B/16 underperforms the ResNet architectures on the reported comparisons.

A Effect of Model Architectures

B Appendix

The appendix checklist covers dataset documentation and access, hosting and long-term preservation, and reproducibility of benchmark results.

B Appendix

  • Dataset submissions should document intended uses, provide reviewer access, and include responsibility and licensing statements.Source: B Appendix · p. 14
  • Dataset accessibility guidance calls for access links, a reading explanation, long-term preservation, and an explicit license.Source: B Appendix · p. 14
  • Structured metadata and persistent identifiers are recommended, while benchmark supplements should make results easily reproducible through accessible datasets, code, and evaluation procedures.Source: B Appendix · pp. 14–15

Checklist

The checklist reports that the paper describes its limitations and provides code and training details, while marking compute reporting and asset licensing as incomplete. Several human-subjects and data-privacy items are marked not applicable.

Checklist

  • The authors report describing the work’s limitations in Section 8 and mark discussion of potential negative societal impacts as not applicable.Source: Checklist · p. 15
  • The checklist marks code and training details as available and reports error bars in Table 3, but says compute resources were not reported.The code URL is listed as https://github.com/aengusl/spawrious.Source: Checklist · p. 15
  • The checklist marks citation of pretrained models and architectures as complete, but says asset licensing is not discussed in the paper and is available in the GitHub repository.Source: Checklist · p. 15
  • The checklist marks data privacy, offensive-content, and human-subjects items as not applicable.Source: Checklist · p. 15

Related papers

Loading 2303.05470v3…