Discovery and Spatial Characterisation of Multiple Shortcut Groups for Auditing Vision Model Bias

Akshit Achara; Vishnunarayan Manickam; Thomas Day; Esther Puyol Anton; Alexander Hammers; Andrew P. King

cs.CVAug 14, 2026

Discovery and Spatial Characterisation of Multiple Shortcut Groups for Auditing Vision Model Bias

Authors: Akshit Achara, Vishnunarayan Manickam, Thomas Day, Esther Puyol Anton, Alexander Hammers, Andrew P. King

Organizations: School of Biomedical Engineering and Imaging Sciences, King’s College London, UK

Abstract

Deep learning models trained on datasets with spurious correlations can achieve high average accuracy whilst relying on shortcut features that do not generalise out of distribution. Whilst out-of-distribution testing highlights subgroup performance disparities arising from shortcut learning, it does not localise the regions within images that are associated with it. Existing research mostly uses attribution maps from interpretability methods to understand the spatial nature of spurious correlations. For example, conditional alignment methods separate task-relevant evidence from evidence tied to spurious correlations by comparing attribution maps from a task model, a sensitive attribute model, and a bias-reduced reference model. This yields shortcut-aligned and task-aligned contribution maps for each image. However, existing methods aggregate these maps across the dataset, potentially masking recurring spatial shortcut patterns that occur only in subsets of images. We address this limitation by grouping per-image shortcut and task contribution maps into recurring spatial patterns using K-means and non-negative matrix factorisation, and visualising the resulting shortcut groups through contribution maps and representative examples. Across CelebA, CheXpert, Waterbirds, Camelyon17, and ISIC2019, and across ResNet and ViT models, the discovered shortcut groups reveal both shared and distinct spatial patterns of shortcut and task contribution, with varying subgroup composition and error rates, enabling targeted inspection of image subsets with higher error rates. We perform input occlusion and internal test-time interventions to show that masking or suppressing task contribution regions substantially degrades the model classification performance and propose a combined shortcut suppression and task amplification feature intervention approach which generally reduces performance disparities.

Figures & tables

Dataset	$(Y{=}0,A{=}0)$	$(Y{=}0,A{=}1)$	$(Y{=}1,A{=}0)$	$(Y{=}1,A{=}1)$
CelebA	500	4500	4500	500
CheXpert	500	4500	4500	500
Waterbirds 2 2 2 We deliberately introduce a higher dataset bias for Waterbirds as the group-level performance disparities were minimal for the model trained using 500 discordant samples.	4900	100	100	4900
Camelyon17	500	14500	14500	500
ISIC2019	2000	500	500	2000

Table 1: Biased training splits with strong label–attribute correlations. Counts are reported for the four

(Y,A)

(label, attribute) groups.

Dataset	Backbone	LPG ( $f_{TS}$ ) $\downarrow$	LPG ( $f_{BA}$ ) $\downarrow$	$f_{SA}$ Acc. $\uparrow$
CelebA	ViT	0.24	0.07	0.96
CelebA	ResNet	0.28	0.05	0.98
CheXpert	ViT	0.37	0.10	0.94
CheXpert	ResNet	0.41	0.12	0.94
Waterbirds	ViT	0.42	0.06	0.97
Waterbirds	ResNet	0.24	0.03	0.97

Table 2: Balanced training substantially reduces the largest performance gap for every dataset and backbone, while sensitive attributes are highly predictable. Values are averaged across independently trained models.

Figure 1: ViT shortcut groups separate recurring spatial patterns with distinct subgroup composition and error rates. Each panel shows a shortcut group prototype and corresponding top $10$ representative images that most strongly express the pattern along with their error rates and subgroup composition.

Figure 2: ResNet shortcut groups also reveal recurring spatial patterns with distinct subgroup composition and error rates. Each panel shows a shortcut group prototype and corresponding top $10$ representative images that most strongly express the pattern along with their error rates and subgroup composition.

Figure 3: Shortcut group risk score derived subsets contain a large fraction of model errors. The dashed line shows the random baseline, under which

20\%

of the total errors are expected to be found, while the orange and blue bars show the percentage of total errors found in the selected top

20\%

of (held-out) images ranked by the shortcut group derived risk score for K-means and NMF, respectively. The error bars indicate the standard deviation across independently trained models (after averaging over 5 random held-out sets for each model).

Dataset	Model	Source	No intervention		Shortcut		Task		Random
			LPG	Acc.	$\Delta\mathrm{LPG}_{\mathrm{red}}$	$\Delta$ Acc.	$\Delta\mathrm{LPG}_{\mathrm{red}}$	$\Delta$ Acc.	$\Delta\mathrm{LPG}_{\mathrm{red}}$	$\Delta$ Acc.
CelebA	ResNet	Shortcut Group	27.6	86.9	-1.2	-.7	-42.2	-10.5	$-10.9{\pm}5.0$	$-2.4{\pm}1.5$
	ResNet	Image	27.6	86.9	-.3	-.2	-34.6	-9.2	$-10.8{\pm}4.5$	$-2.2{\pm}1.5$
	ViT	Shortcut Group	23.7	87.2	-4.4	-.8	-50.7	-12.5	$-8.8{\pm}1.1$	$-1.8{\pm}.2$
	ViT	Image	23.7	87.2	-3.1	-.1	-24.7	-6.2	$-7.1{\pm}2.4$	$-1.5{\pm}.4$
CheXpert	ResNet	Shortcut Group	41.4	76.5	+5.9	+.5	-5.4	-4.0	$+5.7{\pm}8.6$	$-1.5{\pm}1.2$

Table 3: Masking task contribution regions results in severe performance disparities. Across most datasets and models, performance degradation is highest under task-region masking, intermediate under random masking, and lowest under shortcut-region masking. Random masking values are averaged over 5 random masks for each independently trained model and reported with standard deviation across models. Blue denotes the largest improvement and red the highest performance degradation for each dataset–model pair.

		Shuffled		Task		Shortcut Group		Image
Dataset	Backbone	$\Delta\mathrm{LPG}_{\mathrm{red}}$	$\Delta$ Acc.	$\Delta\mathrm{LPG}_{\mathrm{red}}$	$\Delta$ Acc.	$\Delta\mathrm{LPG}_{\mathrm{red}}$	$\Delta$ Acc.	$\Delta\mathrm{LPG}_{\mathrm{red}}$	$\Delta$ Acc.
CelebA	ResNet	$+1.4{\pm}2.3$	$+.1{\pm}.2$	${\color[rgb]{1,0,0}-8.1{\pm}5.0}$	${\color[rgb]{1,0,0}-3.7{\pm}.6}$	${\color[rgb]{0,0,1}+7.7{\pm}5.0}$	${\color[rgb]{0,0,1}+1.6{\pm}.5}$	$+5.3{\pm}1.7$	$+.8{\pm}.4$
CelebA	ViT	$-.3{\pm}.3$	$-.1{\pm}.1$	${\color[rgb]{1,0,0}-11.8{\pm}5.7}$	${\color[rgb]{1,0,0}-4.7{\pm}2.3}$	${\color[rgb]{0,0,1}+4.0{\pm}4.4}$	${\color[rgb]{0,0,1}+1.3{\pm}1.1}$	$+3.1{\pm}2.4$	$+.8{\pm}.6$
CheXpert	ResNet	$-1.7{\pm}1.3$	$-.8{\pm}.5$	${\color[rgb]{1,0,0}-10.5{\pm}5.5}$	${\color[rgb]{1,0,0}-3.6{\pm}1.7}$	$+.1{\pm}4.5$	$+.3{\pm}1.3$	${\color[rgb]{0,0,1}+5.2{\pm}3.9}$	${\color[rgb]{0,0,1}+1.4{\pm}1.3}$
CheXpert	ViT	$0.0{\pm}2.2$	$-1.0{\pm}.3$	${\color[rgb]{1,0,0}-23.0{\pm}8.0}$	${\color[rgb]{1,0,0}-9.7{\pm}1.5}$	${\color[rgb]{0,0,1}+12.8{\pm}7.6}$	${\color[rgb]{0,0,1}+3.2{\pm}.5}$	$+6.5{\pm}1.6$	$+.8{\pm}.1$
Waterbirds	ResNet	$-.9{\pm}1.0$	$-1.0{\pm}.3$	${\color[rgb]{1,0,0}-39.6{\pm}4.7}$	${\color[rgb]{1,0,0}-20.3{\pm}1.4}$	${\color[rgb]{0,0,1}+5.4{\pm}1.5}$	${\color[rgb]{0,0,1}+3.7{\pm}.7}$	$+.6{\pm}.4$	$+.8{\pm}.2$

Table 4: Combined shortcut suppression and task amplification generally reduces subgroup disparities, whereas task suppression degrades performance. Values show changes post-intervention, reported as mean and standard deviation across independently trained models. Shuffled values are averaged over five random shuffles per model. Blue denotes the largest improvement and red the highest performance degradation for each dataset–model pair.

Dataset	Source	Iteration $1$			$\bar{\rho}_{\mathrm{sc}_{0}}$		Iteration $5$
		$\bar{\rho}_{\mathrm{sc}}$	$\Delta\mathrm{LPG}_{\mathrm{red}}$	$\Delta$ Acc.	$\Delta\mathrm{LPG}_{\mathrm{red}}$	$\Delta$ Acc.	$\bar{\rho}_{\mathrm{sc}}$	$\Delta\mathrm{LPG}_{\mathrm{red}}$	$\Delta$ Acc.
CelebA	Shortcut Group	+.110	+4.0	+1.3	+7.8	+1.2	-.025	+9.8	+.1
CelebA	Image	+.112	+3.1	+.8	+2.7	+.7	-.024	+1.1	+.7
CheXpert	Shortcut Group	+.062	+12.8	+3.2	+20.3	-.4	+.003	+20.3	-.4
CheXpert	Image	+.035	+6.5	+.8	+6.5	+.8	-.087	+12.6	+1.9
Waterbirds	Shortcut Group	+.036	+9.7	+3.7	+13.9	+5.2	-.019	+17.7	+4.5

Table 5: Iterative interventions reduce shortcut alignment and can decrease subgroup disparities beyond a single intervention. Values are mean changes relative to no intervention across independently trained models. The

\bar{\rho}_{\mathrm{sc}_{0}}

stage is the last iteration with a positive mean shortcut partial correlation across test images (or closest to

0

if mean across all iterations is negative). For iteration

1

and

5

, we also show the mean shortcut partial correlation

\bar{\rho}_{\mathrm{sc}}

. Blue denotes the largest improvement per dataset.

Appendix figures & tables48 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: Quadrant decomposition of shortcut alignment score using ViT based shortcut groups. Orange shows $|z_{1,i}z_{2,i}|$ within the indicated quadrant.

Figure 5: Quadrant decomposition of shortcut alignment score using ResNet based shortcut groups. Orange shows $|z_{1,i}z_{2,i}|$ within the indicated quadrant.

Figure 6: Quadrant decomposition of task alignment score using ViT based shortcut groups. Blue shows $|z_{1,i}z_{2,i}|$ within the indicated quadrant.

Figure 7: Quadrant decomposition of task alignment score using ResNet based shortcut groups. Blue shows $|z_{1,i}z_{2,i}|$ within the indicated quadrant.

Figure 8: Risk-score-based error identification is robust across grid-partition resolutions. Coloured bars represent the percentage of total errors captured within the top 20% of samples ranked by risk score; the dashed line indicates the 20% random baseline. Error bars denote the standard deviation across independently trained ViT models averaged over five (held-out) splits.

Figure 9: Increasing the number of shortcut groups beyond

K=8

yields comparable error-subset identification. The dashed line marks random selection at the same 20% inspection budget. Values are first averaged across five different held-out sets and the error bars indicate the deviations across independently trained models.

Figure 10: Risk-score-ranked subsets consistently capture higher error proportions across varying inspection budgets. Curves show the percentage of total errors identified as the subset size increases; the dashed line represents the random selection baseline. Error bars denote standard deviations across independently trained models averaged over five evaluation splits.

Dataset	Model	Source	Shortcut		Task		Random
			$\Delta\mathrm{LPG}_{\mathrm{red}}$	$\Delta$ Acc.	$\Delta\mathrm{LPG}_{\mathrm{red}}$	$\Delta$ Acc.	$\Delta\mathrm{LPG}_{\mathrm{red}}$	$\Delta$ Acc.
CelebA	ResNet	Dataset	+2.3	+.9	-34.2	-8.6	$-19.6{\pm}7.4$	$-5.1{\pm}2.3$
		Shortcut Group	-1.1	-.3	-39.4	-10.1	$-17.9{\pm}6.7$	$-5.1{\pm}1.8$
		Image	+1.5	-.2	-32.8	-9.5	$-16.7{\pm}6.9$	$-4.2{\pm}2.1$
	ViT	Dataset	-5.5	-.8	-44.5	-10.3	$-10.4{\pm}5.0$	$-2.3{\pm}.8$
		Shortcut Group	-4.3	-.8	-49.6	-12.3	$-11.8{\pm}3.1$	$-2.8{\pm}.5$

Appendix

Table 10: Masking task regions degrades model performance across all datasets, while masking shortcut regions either results in lower performance degradations or shows slight improvements. Random masking values are averaged over 5 random masks for each independently trained model and reported with standard deviation across models. Blue and red indicate the highest increase (or smallest drop) and largest drop, respectively, for each dataset–model pair.

Dataset	Model	Dataset		Shortcut Group		Image
		$\Delta\mathrm{LPG}_{\mathrm{red}}$	$\Delta$ Acc.	$\Delta\mathrm{LPG}_{\mathrm{red}}$	$\Delta$ Acc.	$\Delta\mathrm{LPG}_{\mathrm{red}}$	$\Delta$ Acc.
CelebA	ResNet	$+4.9{\pm}2.3$	$+1.4{\pm}.2$	${\color[rgb]{0,0,1}+7.7{\pm}5.0}$	${\color[rgb]{0,0,1}+1.6{\pm}.5}$	$+5.3{\pm}1.7$	$+.8{\pm}.4$
CelebA	ViT	$+2.1{\pm}2.8$	$+1.1{\pm}.4$	${\color[rgb]{0,0,1}+4.0{\pm}4.4}$	${\color[rgb]{0,0,1}+1.3{\pm}1.1}$	$+3.1{\pm}2.4$	$+.8{\pm}.6$
CheXpert	ResNet	$-3.0{\pm}4.4$	$-.7{\pm}.5$	$+.1{\pm}4.5$	$+.3{\pm}1.3$	${\color[rgb]{0,0,1}+5.2{\pm}3.9}$	${\color[rgb]{0,0,1}+1.4{\pm}1.3}$
CheXpert	ViT	$+5.4{\pm}4.1$	$+.8{\pm}1.9$	${\color[rgb]{0,0,1}+12.8{\pm}7.6}$	${\color[rgb]{0,0,1}+3.2{\pm}.5}$	$+6.5{\pm}1.6$	$+.8{\pm}.1$
Waterbirds	ResNet	$+2.6{\pm}1.5$	$+2.7{\pm}.6$	${\color[rgb]{0,0,1}+5.4{\pm}1.5}$	${\color[rgb]{0,0,1}+3.7{\pm}.7}$	$+.6{\pm}.4$	$+.8{\pm}.2$

Appendix

Table 11: Shortcut group-level combined interventions result in the strongest improvements in LPG reduction and accuracy across most datasets. Values are the mean change in LPG reduction and accuracy relative to no-intervention with standard deviation across independently trained models.

Dataset	Model	Shortcut Group		Image
		$\Delta\mathrm{LPG}_{\mathrm{red}}$	$\Delta$ Acc.	$\Delta\mathrm{LPG}_{\mathrm{red}}$	$\Delta$ Acc.
Shortcut suppression
CelebA	ResNet	$+5.8{\pm}3.9$	$+1.7{\pm}.4$	$+6.3{\pm}2.7$	$+1.1{\pm}.5$
CelebA	ViT	$+1.4{\pm}4.4$	$+.9{\pm}.9$	$+3.7{\pm}3.1$	$+1.2{\pm}.7$
CheXpert	ResNet	$-1.0{\pm}1.9$	$+.5{\pm}.8$	$0.0{\pm}2.5$	$+.1{\pm}.9$
CheXpert	ViT	$+13.8{\pm}4.2$	$+3.0{\pm}.7$	$+12.7{\pm}4.0$	$+2.0{\pm}.8$

Appendix

Table 12: Intervention stability across independently trained models. Values are mean

\pm

standard deviation changes relative to the corresponding model without intervention.

Table 13: Combined shortcut suppression and task amplification-based internal intervention target layer comparison. Values are mean changes relative to the no-intervention model, averaged over all five datasets across independently trained models.

Intervention target	Dataset		Shortcut Group		Image
	$\Delta\mathrm{LPG}_{\mathrm{red}}$	$\Delta$ Acc.	$\Delta\mathrm{LPG}_{\mathrm{red}}$	$\Delta$ Acc.	$\Delta\mathrm{LPG}_{\mathrm{red}}$	$\Delta$ Acc.
V	+2.9	+.7	+7.4	+1.8	+4.4	+.9
V + MLP	+1.0	+.2	+6.4	+1.4	+1.7	0.0
Q + K + V + MLP	-22.4	-9.2	-19.9	-10.5	-26.1	-11.9
Q + K + V	-16.9	-7.1	-19.8	-9.2	-22.1	-10.4
Q + K	-8.8	-4.3	-6.6	-4.6	-12.2	-6.1

Appendix

Table 14: Combined intervention on value vector results in reduction in disparities and improvements in accuracy. Values are mean changes relative to no intervention, averaged over all five datasets and independently trained models. Blue denotes the largest improvement within each contribution map source and metric.

Dataset	Source	Iteration $1$			$\bar{\rho}_{\mathrm{sc}_{0}}$		Iteration $5$
		$\bar{\rho}_{\mathrm{sc}}$	$\Delta\mathrm{LPG}_{\mathrm{red}}$	$\Delta$ Acc.	$\Delta\mathrm{LPG}_{\mathrm{red}}$	$\Delta$ Acc.	$\bar{\rho}_{\mathrm{sc}}$	$\Delta\mathrm{LPG}_{\mathrm{red}}$	$\Delta$ Acc.
CelebA	Shortcut Group	-.301	+7.7	+1.6	+7.7	+1.6	-.397	+11.3	+1.5
CelebA	Image	-.240	+5.3	+.8	+5.3	+.8	-.352	+6.6	+1.3
CheXpert	Shortcut Group	+.119	+.1	+.3	-6.2	-.8	+.069	-6.2	-.8
CheXpert	Image	+.018	+5.2	+1.4	+5.2	+1.4	-.123	-.8	0.0
Waterbirds	Shortcut Group	+.050	+5.4	+3.7	+5.6	+4.0	+.024	+5.6	+4.0

Appendix

Table 15: Iterative interventions can decrease subgroup disparities beyond a single intervention. Values are mean changes relative to no intervention across independently trained models. The

\bar{\rho}_{\mathrm{sc}_{0}}

stage is the last iteration with a positive mean shortcut partial correlation across test images (or closest to

0

if mean across all iterations is negative). For iteration

1

and

5

, we also show the mean shortcut partial correlation

\bar{\rho}_{\mathrm{sc}}

. Blue denotes the largest improvement per dataset.

Figure 11: Shortcut contributions are suppressed over iterations whereas task contributions are comparatively stable. Iterative contribution maps under shortcut suppression on CheXpert using ViT.

Figure 12: Shortcut contributions are comparatively stable whereas task contributions often vary across iterations. Iterative contribution maps under task amplification on CheXpert using ViT.

Figure 13: Shortcut contributions are suppressed over iterations whereas task contributions often vary. Iterative contribution maps under combined shortcut suppression and task amplification on CheXpert using ViT.

Figure 14: Shortcut contributions are suppressed from the second iteration onwards and the task contributions are comparatively stable. Iterative contribution maps under one initial combined intervention followed by only shortcut suppression on CheXpert using ViT.

Panel of Table 16 — Table 16: Task contributions are concentrated inside bird/lung regions, while shortcut contributions are primarily in external context on aggregate and across shortcut groups. For each shortcut group, we form membership-weighted shortcut group prototypes and annotated masks, then report the percentage of the prototype’s top- $10\%$ regions overlapping the weighted mask. Values are computed from the prototype maps shown in the corresponding visualisations. Values are based on one ViT model per dataset.

Figure 16: CheXpert , ResNet, Grad-CAM: NMF shortcut group contribution maps.

Figure 17: CheXpert , ResNet, Grad-CAM: K-means shortcut group contribution maps.

Figure 18: CheXpert , ResNet, LRP: NMF shortcut group contribution maps.

Figure 19: CheXpert , ResNet, LRP: K-means shortcut group contribution maps.

Figure 20: CheXpert , ViT, AttnLRP: NMF shortcut group contribution maps.

Figure 21: CheXpert , ViT, AttnLRP: K-means shortcut group contribution maps.

Figure 22: CheXpert , ViT, TiS: NMF shortcut group contribution maps.

Figure 23: CheXpert , ViT, TiS: K-means shortcut group contribution maps.

Figure 24: CheXpert , ViT, CX: NMF shortcut group contribution maps.

Figure 25: CheXpert , ViT, CX: K-means shortcut group contribution maps.

Figure 26: Camelyon17 , ResNet, Grad-CAM: NMF shortcut group contribution maps.

Figure 27: Camelyon17 , ResNet, Grad-CAM: K-means shortcut group contribution maps.

Figure 28: Camelyon17 , ResNet, LRP: membership-weighted NMF shortcut group contribution maps.

Figure 29: Camelyon17 , ResNet, LRP: K-means shortcut group contribution maps.

Figure 30: Camelyon17 , ViT, AttnLRP: NMF shortcut group contribution maps.

Figure 31: Camelyon17 , ViT, AttnLRP: K-means shortcut group contribution maps.

Figure 32: Camelyon17 , ViT, TiS: NMF shortcut group contribution maps.

Figure 33: Camelyon17 , ViT, TiS: K-means shortcut group contribution maps.

Figure 34: Camelyon17 , ViT, CX: NMF shortcut group contribution maps.

Figure 35: Camelyon17 , ViT, CX: K-means shortcut group contribution maps.

Figure 36: ISIC2019 , ResNet, Grad-CAM: NMF shortcut group contribution maps.

Figure 37: ISIC2019 , ResNet, Grad-CAM: K-means shortcut group contribution maps.

Figure 38: ISIC2019 , ResNet, LRP: NMF shortcut group contribution maps.

Figure 39: ISIC2019 , ResNet, LRP: K-means shortcut group contribution maps.

Figure 40: ISIC2019 , ViT, AttnLRP: NMF shortcut group contribution maps.

Figure 41: ISIC2019 , ViT, AttnLRP: K-means shortcut group contribution maps.

Figure 42: ISIC2019 , ViT, TiS: NMF shortcut group contribution maps.

Figure 43: ISIC2019 , ViT, TiS: K-means shortcut group contribution maps.

Figure 44: ISIC2019 , ViT, CX: NMF shortcut group contribution maps.

Figure 45: ISIC2019 , ViT, CX: K-means shortcut group contribution maps.

Explore similar work

May 27, 2026

Bias Leaves a Gradient Trail: Label-Free Bias Identification via Gradient Probes on Concept Decompositions

Vision classifiers can exploit spurious correlations, achieving high in-distribution accuracy yet failing under distribution shift. Existing approaches to bias mitigation and analysis often depend on curated datasets, spurious-attribute or group labels, or retraining, which may be infeasible once a model is deployed or the relevant bias is unknown. We present a bias-label-free, post-hoc method for identifying spurious concepts in frozen vision models, relying only on standard class labels from a held-out audit dataset. For each target class, we collect patches from inputs predicted as that class and apply non-negative matrix factorization to intermediate activations to obtain a bank of interpretable concept vectors. Candidate concepts are then ranked with a bias estimator derived from their interaction with backpropagated gradients on misclassified examples: bias concepts tend to get activated when correcting false negatives and suppressed when correcting false positives. On Colored MNIST and Waterbirds the method recovers concepts aligned with the known spurious cue, and on CelebA it surfaces decision-relevant directions that only partially coincide with the annotated gender attribute; suppressing the top-ranked concepts at inference time improves worst-group accuracy by up to 17.9 percentage points on Waterbirds and 10.4 on CelebA without any retraining or parameter updates. Our method identifies decision-relevant spurious directions that need not coincide with annotated ones, providing both an interpretable auditing tool and an actionable debiasing handle for frozen vision models. Code is available at https://github.com/vitryt/label-free-bias-identification.

Thomas Vitry, Kieran Edgeworth, Stefan Wermter +1

University of Hamburg, Hamburg, Germany · Ecole Normale Superieure de Rennes, Rennes, France

Debiased ML Spurious Correlation Robustness

Ask PDF

May 13, 2026

Shortcut Mitigation via Spurious-Positive Samples

Shortcut mitigation strategies commonly rely on training data annotations, group-balanced held-out data or the presence of all groups, i.e., all combinations of (spurious) attributes and classes, in the training data. However, these requirements are rarely met in practice. We instead propose a method for targeted model analysis to identify a small set of instances in which the model relies on spurious attributes. Using that set and following ``this feature should not be used for prediction'' reasoning, we identify highly relevant neurons in an intermediate layer and regularize their impact. This ensures that models learn to depend on informative features rather than being right for the wrong reasons, thereby improving robustness without requiring additional balanced held-out data or annotations.

Phuong Quynh Le, Jörg Schlötterer, Sari Sadiya +2

University of Marburg · Goethe University Frankfurt

Neural Network Robustness Spurious Correlation Robustness Shortcut Learning

Ask PDF

Feb 17, 2026

Mitigating Shortcut Learning via Feature Disentanglement in Medical Imaging: A Benchmark Study

Although deep learning models in medical imaging often achieve excellent classification performance, they can rely on shortcut learning, exploiting spurious correlations or confounding factors that are not causally related to the target task. This poses risks in clinical settings, where models must generalize across institutions, populations, and acquisition conditions. Feature disentanglement is a promising approach to mitigate shortcut learning by separating task-relevant information from confounder-related features in latent representations. In this study, we systematically evaluated feature disentanglement methods for mitigating shortcuts in medical imaging, including adversarial learning and latent space splitting based on dependence minimization. We assessed classification performance and disentanglement quality using latent space analyses across one artificial and two medical datasets with natural and synthetic confounders. We also examined robustness under varying levels of confounding and compared computational efficiency across methods. We found that shortcut mitigation methods improved classification performance under strong spurious correlations during training. Latent space analyses revealed differences in representation quality not captured by classification metrics, highlighting the strengths and limitations of each method. Model reliance on shortcuts depended on the degree of confounding in the training data. The best-performing models combine data-centric rebalancing with model-centric disentanglement, achieving stronger and more robust shortcut mitigation than rebalancing alone while maintaining similar computational efficiency. The project code is publicly available at https://github.com/berenslab/medical-shortcut-mitigation.

Sarah Müller, Philipp Berens

Hertie Institute for AI in Brain Health, Faculty of Medicine, University of Tübingen, Germany · Tübingen AI Center, University of Tübingen, Germany

Shortcut Learning Disentangled Representation Learning Medical Image Classification Adversarial Representation Learning

Ask PDF