uniform-convergence

This paper establishes a theoretical framework showing that smooth activations in deep neural networks can mitigate the curse of dimensionality in uniform convergence, providing non-asymptotic guarantees and outperforming ReLU networks in worst-case reliability.