Low-Frequency Shortcuts in Texture-Driven Visual Learning
Texture-driven models rely too heavily on a few low-frequency components (LFCs) for decisions, creating 'low-frequency shortcuts.' The study showed that pruning these LFCs shifts model behavior toward higher frequencies, boosting accuracy by up to 8% and improving robustness against low-frequency corruptions. However, this introduces a trade-off when dealing with high-frequency corruptions.