Low-Frequency Shortcuts in Texture-Driven Visual Learning
Utku Şirin, Cathy Hou, David Alvarez-Melis, Stratos Idreos · 2026
Read Low-Frequency Shortcuts in Texture-Driven Visual Learning by Utku Şirin, Cathy Hou, David Alvarez-Melis, Stratos Idreos (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-03493.txt.
Description
Şirin and co-authors show that texture-driven image models lean on a narrow shortcut. Neural networks often learn features that work on training data but fail on new data. Past studies used shape-driven benchmarks. Many real domains are texture-driven. The team compares shortcut behavior in those domains with a standard benchmark.
Texture models base most decisions on a few low-frequency components. Higher-frequency parts carry more predictive power, yet the models largely ignore them. Removing the low-frequency parts from training and test data fixes much of this. In-distribution accuracy rises by up to 10 percent, and out-of-distribution accuracy by up to 40 percent. Large foundation models show the same shortcut, at high compute cost. Small models trained from scratch, once pruned of the shortcut, can beat them. Lower image resolution makes the shortcut worse. The results guide work in new texture-heavy domains.
Catalog record
| CREATOR | Utku Şirin, Cathy Hou, David Alvarez-Melis, Stratos Idreos |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 69 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.03493