Reproduce “Explaining the Saliency Map Sparsity of Adversarially-Trained Neural Networks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yannick Lunk, Atell Yehor Krasnopolsky, Damien Garreau, Leon Bungert
The paper explains how TV regularization can favor sparse-looking input-gradient saliency maps. In a limiting analysis of binary two-layer ReLU classifiers, an ℓ∞ perturbation budget induces an ℓ1 gradient penalty, favoring concentrated gradients. In a two-class CIFAR-10 experiment, thresholded sparsity at 0.02 was 0.92 for TV, 0.31 for adversarial training, and 0.01 naturally; this does not establish the mechanism for arbitrary deep networks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.