Rethinking Noise-Robust Training for Frozen Vision Foundation Models: A Cross-Dataset Benchmark with a Case Study of Small-Loss Failure
A new benchmark study on noise-robust training for frozen vision foundation models reveals that no single method consistently outperforms others across various medical imaging datasets and noise conditions. The research highlights that the choice of method significantly impacts performance, especially with increasing noise severity. Findings suggest that selecting an appropriate method based on the specific noise regime is more crucial than searching for a universally dominant algorithm. AI
IMPACT Highlights the complexity of choosing noise-robust training methods for vision models, suggesting a need for regime-aware selection over a single best algorithm.