IC-742ResNet-50-BN on Waterbirds relies on background as a spurious feature for classification, and this shortcut is invisible to entropy-based confidence metrics
On the Waterbirds benchmark, where background is strongly correlated with class in training but anti-correlated in test, the pre-trained ResNet-50-BN exhibits a shortcut: its predictions are driven by background rather than the bird. The paper shows that within the worst-performing group, samples with entropy below the first quartile have lower prediction accuracy than samples in higher entropy intervals, meaning the most 'confident' predictions are the most likely to be wrong. Grad-CAM visualizations confirm that wrong low-entropy predictions focus on the background while correct ones focus on the bird. Applying entropy-based TTA (TENT) to this model degrades worst-group accuracy from 64.90% to 54.14%, below the pre-adaptation level.
Evidence
correlational
Key metric
Worst-group accuracy on Waterbirds: ResNet-50-BN pre-adaptation 64.90%, TENT 54.14%, MEMO 50.47%, EATA 52.38%, SAR 53.41%; entropy below first quartile in worst group shows lowest accuracy (Fig. 2a)
Caveat
The analysis is on a single architecture (ResNet-50-BN) and a single benchmark (Waterbirds) with binary classification; the Grad-CAM evidence is illustrative rather than exhaustive.