Localising Shortcut Learning in Pixel Space via Ordinal Scoring Correlations for Attribution Representations (OSCAR)

2025-12-21

AI Summary

Deep learning models are prone to shortcut learning—exploiting spurious correlations between non-semantic features and labels or sensitive attributes—leading to biased predictions, particularly in domains like medical imaging where human interpretability is limited and quantitative, pixel-level diagnostic tools are lacking. To address this, we propose the first model-agnostic framework for quantifying and localizing shortcut learning at the pixel level. Our method constructs a dataset-level regional ranking spectrum from attribution maps and employs ordinal correlation analysis, partial correlation, and bias correlation metrics across three complementary models (BA, TS, SA). We further introduce a test-time shortcut region attenuation strategy. Extensive experiments on CelebA, CheXpert, and ADNI demonstrate superior stability, sensitivity, and spatial localization accuracy compared to baselines, while significantly narrowing performance gaps across worst-case subgroups.

Technology Category

• Computer Vision: Bias, Fairness & Privacy
• Machine Learning: Calibration & Uncertainty Quantification
• Natural Language Processing: Interpretability, Analysis, and Evaluation of NLP Models

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Deep neural networks often exploit shortcuts. These are spurious cues which are associated with output labels in the training data but are unrelated to task semantics. When the shortcut features are associated with sensitive attributes, shortcut learning can lead to biased model performance. Existing methods for localising and understanding shortcut learning are mostly based upon qualitative, image-level inspection and assume cues are human-visible, limiting their use in domains such as medical imaging. We introduce OSCAR (Ordinal Scoring Correlations for Attribution Representations), a model-agnostic framework for quantifying shortcut learning and localising shortcut features. OSCAR converts image-level task attribution maps into dataset-level rank profiles of image regions and compares them across three models: a balanced baseline model (BA), a test model (TS), and a sensitive attribute predictor (SA). By computing pairwise, partial, and deviation-based correlations on these rank profiles, we produce a set of quantitative metrics that characterise the degree of shortcut reliance for TS, together with a ranking of image-level regions that contribute most to it. Experiments on CelebA, CheXpert, and ADNI show that our correlations are (i) stable across seeds and partitions, (ii) sensitive to the level of association between shortcut features and output labels in the training data, and (iii) able to distinguish localised from diffuse shortcut features. As an illustration of the utility of our method, we show how worst-group performance disparities can be reduced using a simple test-time attenuation approach based on the identified shortcut regions. OSCAR provides a lightweight, pixel-space audit that yields statistical decision rules and spatial maps, enabling users to test, localise, and mitigate shortcut reliance. The code is available at https://github.com/acharaakshit/oscar

Problem

• Quantifies shortcut learning in deep neural networks
• Localizes shortcut features using attribution map correlations
• Provides pixel-space audit to test and mitigate model bias

Innovation

• Model-agnostic framework quantifying shortcut learning via attribution maps
• Correlates rank profiles across baseline, test, and attribute predictor models
• Produces metrics and spatial maps to localize and mitigate shortcut features