Uniformly Stable Algorithms for Adversarial Training and Beyond

Authors: Jiancong Xiao, Jiawei Zhang, Zhi-Quan Luo, Asuman E. Ozdaglar

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In practical scenarios, we demonstrate the efficacy of ME-A in mitigating the issue of robust overfitting. ... 6. Experiments
Researcher Affiliation Academia 1University of Pennsylvania, PA, USA; 2Massachusetts Institute of Technology, MA, USA; 3The Chinese University of Hong Kong, Shenzhen, China.
Pseudocode Yes Algorithm 1 Moreau Envelope-A
Open Source Code Yes 2Code is publicly available at https://github.com/Jiancong Xiao/Moreau-Envelope-SGD.
Open Datasets Yes It can be observed in experiments on common datasets such as SVHN, CIFAR-10/100. ... Carmon et al. (2019)
Dataset Splits No The paper mentions using CIFAR-10, SVHN, and CIFAR-100 datasets and discusses training and test accuracy, but it does not explicitly provide the training/validation/test splits (e.g., percentages or sample counts) or state that standard splits were used.
Hardware Specification No The paper does not explicitly state any specific hardware details such as GPU models, CPU types, or memory used for running the experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers used for replicating the experiments.
Experiment Setup Yes Weight decay is set to be 5 × 10−4. Based on Theorem 4.7, the step size αt of updating u is set to be 1/pt, then τt = (t − 1)/t. ... For the attack algorithms, we use ϵ = 8/255. The attack step size is set to be ϵ/4. We use piece-wise learning rates, which are equal to 0.1, 0.01, 0.001 for epochs 1 to 100, 101 to 150, and 151 to 200, respectively.