Low-Frequency Shortcuts in Texture-Driven Visual Learning
Researchers have identified a new type of shortcut learning in visual AI models that are trained on texture-driven datasets. These models tend to rely on low-frequency components for classification, even when the crucial information lies in finer details. By pruning these low-frequency components, the models' accuracy on in-distribution data improved by up to 8% and their robustness to certain corruptions increased significantly. AI
IMPACT Identifies a new vulnerability in AI models, potentially impacting their reliability in real-world texture-heavy applications.