Why noise-robust training fails for frozen vision models

Frozen vision foundation models are popular in medical imaging, but denoising methods designed for end-to-end training don't work reliably when the backbone is frozen. Researchers benchmarked eight denoising approaches across 150 conditions (five medical datasets, multiple backbones, varying noise levels) and found no universal winner—performance gaps reach 18.8 percentage points depending on which method you pick. The culprit: the standard assumption that clean samples have lower loss than noisy ones collapses under asymmetric noise, where loss distributions overlap by over half. They propose a regime-aware selector to match methods to specific noise patterns rather than hunting for a single algorithm.