Research map / Label-efficient and robust learning / OOD and anomaly detection

Do ImageNet-trained Models Learn Shortcuts? The Impact of Frequency Shortcuts on Generalization

CVPR 2025 · Label-efficient and robust learning › OOD and anomaly detection · approach: Robust fine-tuning for OOD generalization · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes HFSS, a hierarchical random-search method to efficiently find Fourier-domain frequency shortcuts in large-scale ImageNet-trained CNN and transformer models, and links them to OOD generalization.

Problem

Identifying frequency shortcuts with prior methods is computationally infeasible for models trained on large datasets like ImageNet-1k, leaving unclear how such shortcuts affect generalization under OOD shifts.

Approach

A hierarchical, coarse-to-fine random sampling of frequency patches across search stages, scoring joint frequency subsets by class-wise loss, to find Dominant Frequency Maps (DFMs); shortcut classes are flagged by TPR on DFM-filtered images, and performance is compared across ID, corruption, rendition, and adversarial OOD tests.

Key results

• HFSS reduces search time on ImageNet-1k from an estimated 8500 h (354 days) to 220 h (9.2 days) and on CIFAR-10 from 7.5 h to 0.5 h relative to Wang et al.
• On ImageNet-10, HFSS finds strong frequency shortcuts (TPR>=0.6) for 6 classes versus 2 for the prior method.
• Shortcut-reliant models perform better on IN-v2 and IN-C and under FGSM, but worse on rendition-based IN-R (200 classes); CCT with the strongest shortcut learning achieves 57.73% on IN-C vs 48.85% for ResNet50.

Measured effects reported in the paper

Surprising findings

• Frequency shortcuts help robustness on corruption and adversarial tests but impair generalization to rendition-based OOD sets such as IN-R.
• CCT shows the highest degree of shortcut learning yet also the best adversarial robustness and strong OOD performance, rather than being purely harmed by shortcuts.

Limitations

• Search remains computationally intensive for ImageNet; efficient configuration trades off effectiveness at high thresholds.
• Analysis is restricted to four architectures and ImageNet-trained models; frequency shortcuts may not arise from fixed frequency subsets.
• Benchmark conclusions depend on the chosen TPR threshold and DFM resolution (about 5% of frequencies).

Experimental setup

Datasets: ImageNet-1k CIFAR-10 ImageNet-10 ImageNet-v2 ImageNet-C ImageNet-R ImageNet-Sketch ImageNet-SCT

Benchmarks: ImageNet-1k ImageNet-v2 ImageNet-C ImageNet-R ImageNet-Sketch FGSM adversarial attack on ImageNet-1k ImageNet-SCT ImageNet-10

Metrics: true positive rate (TPR) average TPR of shortcut vs non-shortcut classes search time loss

Baselines: Wang et al. 2023 frequency shortcut method

Builds on: Wang et al. frequency shortcut identification (DFM) simplicity bias ImageNet-SCT

Closest papers by meaning

• Improved Training Technique for Shortcut Models (NeurIPS 2025 · similarity 0.89)
• On the Foundations of Shortcut Learning (ICLR 2024 · similarity 0.88)
• Towards Combating Frequency Simplicity-biased Learning for Domain Generalization (NeurIPS 2024 · similarity 0.87)
• One Step Diffusion via Shortcut Models (ICLR 2025 · similarity 0.86)
• FreSh: Frequency Shifting for Accelerated Neural Representation Learning (ICLR 2025 · similarity 0.85)
• In Search of Forgotten Domain Generalization (ICLR 2025 · similarity 0.85)