On Cross-Validation for Hyperparameter Optimization of Deep Learning Image Classifiers
Not specified in the provided content
Abstract
This paper investigates the effectiveness of different hyperparameter optimization protocols for deep learning image classifiers, particularly in small sample size scenarios.
Reality Card
Cross-validation-based hyperparameter optimization is recommended for small-sample medical image classification as it provides a more reliable estimate of test performance compared to fixed and reshuffled holdout protocols.
Cross-validation showed reductions in absolute performance-estimation error (AEE) that were largest at small sample sizes.
The study's findings may not be generalizable beyond the specific datasets and architectures tested.
Paper to code
Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.