Title: Exploiting Low-Frequency Shortcuts in Texture-Based Visual Learning

Abstract:

Neural networks are prone to shortcut learning, a phenomenon where models acquire features that perform well on training data but fail to generalize to both in-distribution (ID) and out-of-distribution (OOD) test sets. While prior research has predominantly relied on shape-driven standard benchmarks, many real-world application domains are texture-driven. This study investigates shortcut learning within texture-centric fields and contrasts it with findings from conventional benchmarks.

Our analysis reveals that texture-driven domains are heavily influenced by low-frequency shortcuts. Despite the fact that critical classification details reside in high-frequency, fine-grained patterns, these models rely predominantly on a limited number of low-frequency components (LFCs) characterized by skewed spectral behavior. By removing LFCs from both training and testing datasets, we eliminate these shortcuts and establish a more balanced spectral profile. This intervention boosts ID accuracy by as much as 8%.

Furthermore, we demonstrate that these low-frequency shortcuts render models exceptionally susceptible to OOD corruptions, causing accuracy to plummet by up to 70% relative to ID performance. While pruning LFCs enhances robustness against low-frequency disturbances by up to 40%, it creates a trade-off regarding high-frequency corruptions. Specifically, the resulting balanced spectral behavior aids generalization, whereas the heightened reliance on high-frequency features detracts from it. Consequently, OOD accuracy is determined by the complex interplay between these opposing factors.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC