Upper Bounds on the Generalization Error of Deep Learning Models via Local Robustness and Stability

The article presents a novel generalization bound for deep learning models that improves upon existing robustness-based bounds by scaling the robustness term according to the stability of samples within input sub-regions. This approach yields tighter upper bounds on true error rates, addressing the vacuousness often seen in practical applications. Experiments conducted on ImageNet-trained models demonstrate that the proposed bounds provide accurate estimates of generalization performance, making them valuable for practitioners in safety-critical AI applications.