PaperlayerView exact revision on arXiv

Source-linked AI summary

Spread Spurious Attribute: Improving Worst-group Accuracy with Spurious Attribute Estimation

Junhyun Nam, Jaehyung Kim, Jaeho Lee, Jinwoo Shin

arXiv:2204.02070v1cs.LGcs.CV

TL;DR

Spurious correlations can drive low accuracy on underrepresented groups, while worst-group loss minimization requires group supervision. SSA estimates spurious attributes from labeled and unlabeled samples, then uses pseudo-labels to train a robust model; across benchmarks, it outperforms equally annotated baselines and is comparable to fully supervised Group DRO with less than 1.5% annotation.

  • Problem

    Spurious correlations can lead to low test accuracy on underrepresented groups, while worst-group loss minimization requires group supervision for training samples.

    Source: §1 · p. 1

  • Method

    SSA predicts spurious attributes using group-labeled and group-unlabeled samples, then uses those predictions as pseudo-labels to train a robust model with worst-group loss minimization.

    Source: §1 · p. 2

  • Results

    Across benchmarks, SSA consistently outperforms equally annotated baselines and is comparable to fully supervised Group DRO using less than 1.5% (as little as 0.6%) of group-labeled samples.

    Source: §1 · p. 2, §5.1 · p. 6

  • Takeaways & Limitations

    SSA remains dependent on some spurious-attribute annotation, which the authors identify as an important future direction.

    Source: §6 · p. 9

    Worst-group performance should not be taken as a measure of justice, since fair decisions at the group level do not establish individual fairness.

    Source: ETHICS STATEMENT · p. 10

Abstract

from arXiv · show

The paradigm of worst-group loss minimization has shown its promise in avoiding to learn spurious correlations, but requires costly additional supervision on spurious attributes. To resolve this, recent works focus on developing weaker forms of supervision -- e.g., hyperparameters discovered with a small number of validation samples with spurious attribute annotation -- but none of the methods retain comparable performance to methods using full supervision on the spurious attribute. In this paper, instead of searching for weaker supervisions, we ask: Given access to a fixed number of samples with spurious attribute annotations, what is the best achievable worst-group loss if we "fully exploit" them? To this end, we propose a pseudo-attribute-based algorithm, coined Spread Spurious Attribute (SSA), for improving the worst-group accuracy. In particular, we leverage samples both with and without spurious attribute annotations to train a model to predict the spurious attribute, then use the pseudo-attribute predicted by the trained model as supervision on the spurious attribute to train a new robust model having minimal worst-group loss. Our experiments on various benchmark datasets show that our algorithm consistently outperforms the baseline methods using the same number of validation samples with spurious attribute annotations. We also demonstrate that the proposed SSA can achieve comparable performances to methods using full (100%) spurious attribute supervision, by using a much smaller number of annotated samples -- from 0.6% and up to 1.5%, depending on the dataset.

1 INTRODUCTION

Spurious correlations can lead models to rely on shortcuts, hurting accuracy on underrepresented groups, while full worst-group training requires costly group annotations. SSA uses a small annotated set more actively to approach full-supervision performance.

Introduction

  • Spurious correlations can make models learn shortcuts instead of intended solutions, leading to very low test accuracy on underrepresented groups.For example, a cow-versus-camel classifier may rely on grassland versus desert backgrounds rather than the animals.Source: §1 · p. 1
  • Worst-group minimization can improve performance across majority and minority groups, but requires group identity annotations that are costly to collect.The annotations must identify both target and spurious attributes.Source: §1 · p. 1
  • Prior methods use a small group-labeled set for hyperparameter tuning, remain sensitive to those choices, and do not match fully annotated methods.This motivates using group-labeled samples more actively than for tuning alone.Source: §1 · p. 2

Contribution

  • SSA has two phases: pseudo-labeling uses labeled and unlabeled samples to predict spurious attributes, and robust training uses those predictions to train a low-worst-group-loss model.Group-wise adaptive thresholds mitigate pseudo-label bias toward the majority group.Source: §1 · p. 2
  • Using less than 1.5% of group-labeled samples (as little as 0.6%), SSA achieves worst-group accuracy similar to 100% usage on CelebA and MultiNLI.The reported counts are 993 of 182637 on CelebA and 4123 of 288637 on MultiNLI.Source: §1 · p. 2
  • SSA consistently outperforms baseline methods using the same number of group-labeled samples and achieves performance comparable to fully supervised Group DRO across benchmark datasets.The paper also reports that SSA can outperform fully supervised Group DRO on CelebA.Source: §1 · p. 2

2 RELATED WORKS

Prior work addresses worst-group performance with full group annotations or with methods that reduce annotation needs. SSA differs by using group-labeled samples for both training and validation.

Improving worst-group accuracy with group annotations

  • Class-imbalance methods use reweighted losses or resampling when class labels define groups, whereas this paper studies groups combining target and spurious attributes.In the latter setting, complete group identity annotations may not be available.Source: §2 · p. 2

Improving worst-group accuracy without group annotations

  • CNC refers here only to supervised contrastive learning with contrastive batch sampling, not the full process that estimates spurious attributes via ERM..Source: §2 · p. 2
  • Prior annotation-efficient methods identify likely minority samples and upweight them, but still rely on group-labeled samples for sensitive hyperparameter tuning.The paper contrasts its approach with using those samples more actively; Figure 1 depicts SSA using them for training and validation.Source: §2 · p. 3, Figure 1 · p. 3

3 PROBLEM SETUP

The setup seeks a classifier with low worst-group test loss, where groups pair target and spurious attributes. It assumes many samples lack spurious-attribute labels, while a smaller set has them.

Problem setup

  • The assumed setting has no spurious-attribute annotations for the group-unlabeled training set, but provides a separate set with both target and spurious labels..Source: §3 · p. 3
  • The learner aims to minimize worst-group expected test loss over groups defined by target and spurious attribute pairs.The group-conditioned distribution is deterministic on those attributes and stochastic only on the input.Source: §3 · p. 3
  • Training data comprise n group-unlabeled samples with inputs and target labels and m group-labeled samples that also include spurious-attribute labels.The setup assumes group-labeled samples are much costlier to collect, so n ≫ m.Source: §3 · p. 3

4 SPREAD SPURIOUS ATTRIBUTE

SSA estimates spurious attributes from group-labeled and group-unlabeled samples, balances pseudo-label acceptance across groups, and uses the resulting labels for worst-group robust training.

4 SPREAD SPURIOUS ATTRIBUTE

  • SSA has pseudo-labeling and robust-training phases, with Group DRO as the default robust-training method and other robust methods also supported.The generated pseudo-labels supervise target prediction under a small worst-group loss objective.Source: §4 · p. 4

4.1 PSEUDO-LABELING USING GROUP-LABELED AND GROUP-UNLABELED SAMPLES

  • The spurious-attribute predictor trains on both group-labeled and group-unlabeled data, using supervised loss on labeled samples and confidence-thresholded pseudo-label loss on unlabeled samples.The unlabeled set is split so one portion trains the predictor and the other is used for prediction; the paper reports this split benefits performance.Source: §4.1 · p. 4

4.2 BALANCING GENERATED PSEUDO-LABELS VIA ADAPTIVE THRESHOLDS

  • Group-wise adaptive thresholds address majority-group bias by selecting confidence thresholds so that equal numbers of pseudo-group samples contribute to training.The method fixes a threshold for the smallest group, then sets thresholds for other groups to match its accepted sample count.Source: §4.2 · pp. 4–5

4.3 WORST-GROUP LOSS MINIMIZATION WITH ESTIMATED PSEUDO-GROUPS

  • After predictor training, SSA assigns pseudo-labels to all training samples without a threshold, then uses the pseudo-labeled set for worst-group minimization, with group-labeled data for validation.Group DRO is the default, and the framework can also use CNC.Source: §4.3 · p. 5

5 EXPERIMENTS

Across image and language benchmarks, SSA outperforms methods using the same amount of spurious-attribute annotation and is comparable to fully supervised Group DRO. The experiments also examine reduced annotation, pseudo-labeling thresholds, combining SSA with contrastive learning, and class-imbalanced semi-supervised learning.

5.1 MAIN RESULTS

  • Across the image and language benchmarks, SSA outperforms other methods using the same amount of spurious-attribute annotation and matches full-supervision Group DRO, even outperforming it on CelebA.Table 1 covers Waterbirds and CelebA; Table 2 covers MultiNLI and CivilComments-WILDS.Source: §5.1 · p. 6, Table 1 · p. 6, Table 2 · p. 6
  • SSA maintains high worst-group accuracy with 10% of the default validation set; on Waterbirds, that uses 58 annotated samples, including 6 from the worst group.The reported sample count is for training the spurious-attribute predictor.Source: §5.1 · p. 7, Table 3 · p. 7

5.2 DETAILED ANALYSIS ON THE PSEUDO-LABELING PHASE OF SSA

  • On CelebA, group-wise thresholds boost worst-group spurious-attribute prediction accuracy by 7% at both 5% and 10% validation labels; fixed-threshold pseudo-labeling can slightly hurt at 5%.With 5% validation data, the adaptive threshold raises the male-blond pseudo-labeled fraction from 0.68% to 0.88%, close to the population level above 0.8%.Source: §5.2 · pp. 7–8
  • With 5% of CelebA’s validation split, adaptive thresholding brings the male-blond pseudo-labeled fraction to 0.88%, close to its population fraction above 0.8%, versus 0.68% with a fixed threshold.The adaptive-threshold approach also uses relatively uniform numbers of samples from each group to train the spurious-attribute predictor.Source: §5.2 · p. 8, Table 5 · p. 8

5.3 SSA COMBINED WITH SUPERVISED CONTRASTIVE LEARNING

  • SSA combined with either Group DRO or CNC achieves performance comparable to fully supervised counterparts, while neither robust-training method consistently outperforms the other.CNC slightly exceeds Group DRO on Waterbirds, whereas Group DRO does better on CelebA.Source: §5.3 · p. 8, Table 6 · p. 8

5.4 APPLICATION TO GENERAL SEMI-SUPERVISED LEARNING UNDER CLASS IMBALANCE

  • On imbalanced CIFAR-10 semi-supervised tasks with limited labels and differing labeled/unlabeled class distributions, adaptive thresholds reduce the performance gap to the oracle.DARP and CReST gains over the chosen baselines are limited under these conditions, which violate their respective assumptions.Source: §5.4 · p. 9, Table 7 · p. 9

6 CONCLUSION

SSA estimates training-sample group identities from a small set of spurious-attribute annotations, then uses them to train a robust model. The authors note that supervision is still needed and that worst-group performance is not itself a measure of justice.

6 CONCLUSION

  • SSA uses a small set of spurious-attribute-annotated samples to estimate group identities across the training set and train a robust model with existing worst-case loss minimization algorithms.The authors describe the method as effective with limited annotations, but say it does not completely remove the need for spurious-attribute supervision.Source: §6 · p. 9
  • Worst-group performance reports should not be naively treated as a measure of justice, because group-based evaluation does not establish individual fairness.The authors identify this as a possible pitfall of targeting demographic groups such as gender or race.Source: ETHICS STATEMENT · p. 10
  • The authors provide experimental setup and implementation details in appendices A.2, A.3, and A.5, and include source code in supplementary materials.Source: REPRODUCIBILITY STATEMENT · p. 10

A.1 SPREAD SPURIOUS ATTRIBUTE PSEUDOCODE

The appendix points readers to pseudocode for combining Spread Spurious Attribute with Group DRO.

A.1 SPREAD SPURIOUS ATTRIBUTE PSEUDOCODE

  • Algorithm 1 gives pseudocode for Spread Spurious Attribute combined with Group DRO.Source: §A.1 · p. 13

A.2 DISCUSSIONS ON SPLITTING THE GROUP-LABELED AND GROUP-UNLABELED SETS

SSA splits group-unlabeled data so a sample used for final prediction is separated from samples used to train the pseudo-labeler. The appendix describes this as a way to mitigate self-confirmation while limiting the computation required for splitting.

A.2 DISCUSSIONS ON SPLITTING THE GROUP-LABELED AND GROUP-UNLABELED SETS

  • For a particular unlabeled sample, the best prediction split would use the largest possible number of other group-unlabeled samples to train the model.Source: §A.2 · p. 13
  • Training a separate model for every unlabeled sample is computationally infeasible, so SSA instead partitions the unlabeled set into K equally sized subsets and runs the algorithm K times.Source: §A.2 · p. 13
  • The experiments use K = 3, and the authors state that SSA performance may improve with a larger K.Source: §A.2 · p. 13
  • SSA divides the group-labeled set into equal training and validation subsets and keeps that split fixed throughout pseudo-labeling.Source: §A.2 · p. 14
  • Separating pseudo-labeler training samples from samples receiving final predictions helps mitigate self-confirmation, which can reinforce confident predictions under severe group imbalance.The authors report that splitting improves downstream worst-group performance.Source: §A.2 · p. 14

A.3 TRAINING DETAILS

Training details distinguish pseudo-labeling from robust training, with separate tuning criteria for each phase. The appendix also reports the model families and dataset-specific configurations used.

A.3 TRAINING DETAILS

  • Image experiments use pretrained ResNet-50 and natural-language experiments use BERT; pseudo-labeling hyperparameters are tuned by worst-group spurious-label accuracy on the labeled validation subset.The smallest-population group threshold is fixed at 0.95.Source: §A.3 · p. 14
  • For Waterbirds and CelebA pseudo-labeling, SSA tunes learning rate and L2 regularization, uses SGD with momentum 0.9 and batch size 64, and trains for 1k and 45k iterations, respectively.Source: §A.3 · p. 14
  • For MultiNLI and CivilComments-WILDS, SSA follows Liu et al. (2021) without hyperparameter tuning; pseudo-labeling runs for 30k iterations on both datasets.The stated settings include batch size 32 for MultiNLI and 8 for CivilComments-WILDS.Source: §A.3 · pp. 14–15
  • Robust-model hyperparameters are selected to maximize worst-group validation accuracy, while EIIL baseline implementations use dataset-specific procedures.Robust training runs for 300 epochs on Waterbirds and CelebA, 10 on MultiNLI, and 5 on CivilComments; EIIL procedures differ across datasets.Source: §A.3 · p. 15

A.4 RUNTIME ANALYSIS

SSA adds a pseudo-labeling phase before robust training; with K = 3, its overhead is x0.5 on CelebA and at most x4 on Waterbird compared with vanilla Group DRO.

A.4 RUNTIME ANALYSIS

  • With K = 3, SSA’s overhead is x0.5 on CelebA and at most x4 on Waterbird compared with vanilla Group DRO.Runtime was measured on a single Nvidia Titan XP; other methods addressing incomplete group annotation also require additional computation.Source: §A.4 · p. 15
  • SSA requires an additional pseudo-labeling phase compared with vanilla Group DRO.The reported runtime separates pseudo-labeling from robust training.Source: §A.4 · p. 15
  • The pseudo-labeling iteration count was not optimized, leaving room to improve the runtime-performance tradeoff through further search.Source: §A.4 · p. 15

A.5 DETAILS ON SUPERVISED CONTRASTIVE LEARNING

The appendix describes a supervised contrastive loss for separating groups by target and spurious attribute, then specifies an imbalanced semi-supervised classification setup and its evaluation metrics.

A.5 DETAILS ON SUPERVISED CONTRASTIVE LEARNING

  • The contrastive loss uses same-target, different-spurious-attribute samples as positives and different-target, same-spurious-attribute samples as negatives.It is used in addition to standard cross-entropy loss.Source: §A.5 · p. 16
  • For Waterbirds and CelebA, the contrastive setup uses M = 16, with dataset-specific temperatures, weights, optimizers, and learning rates.Both settings use gradient accumulation to update model parameters every 32 batches.Source: §A.5 · p. 16

A.6 DETAILS ON CLASS-IMBALANCED SEMI-SUPERVISED LEARNING

  • The imbalanced classification setup defines labeled pairs and unlabeled inputs, orders labeled class counts from largest to smallest, and sets unlabeled classes to equal sizes.The labeled imbalance ratio is the largest-to-smallest class sample-size ratio, with class counts set using an exponential decay function.Source: §A.6 · p. 16
  • Performance under class imbalance is reported using balanced accuracy and geometric mean scores, averaged across three random trials with standard deviations.Both metrics summarize class-wise sensitivity, using arithmetic and geometric means, respectively.Source: §A.6 · p. 16

Related papers

Loading 2204.02070v1…