Nonrobust features in biomedical images improve in-distribution accuracy on MedMNIST tasks but degrade performance on shifted data like MedMNIST-C, while robust models show the opposite pattern.
This is expected, as larger perturbations impose a stronger constraint on the features that can be reliably used for classification by the model
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
eess.IV 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Useful nonrobust features are ubiquitous in biomedical images
Nonrobust features in biomedical images improve in-distribution accuracy on MedMNIST tasks but degrade performance on shifted data like MedMNIST-C, while robust models show the opposite pattern.