Zeroth-Order Hard-Thresholding: Gradient Error vs. Expansivity

Authors: William de Vazelhes, Hualin Zhang, Huimin Wu, Xiaotong Yuan, Bin Gu

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5 Experiments
Researcher Affiliation Academia Mohamed bin Zayed University of Artificial Intelligence Nanjing University of Information Science & Technology
Pseudocode Yes Algorithm 1: Stochastic Zeroth-Order Hard-Thresholding (SZOHT)
Open Source Code Yes Our code will be included in the supplementary material
Open Datasets Yes We use three datasets: port3, port4 and port5 from the OR-library [3], of respective dimensions d = 89; 98; 225. We consider three different datasets for the attacks: MNIST, CIFAR, and Imagenet, of dimension respectively d = 784; 3072; 268203.
Dataset Splits No The paper mentions using datasets like MNIST, CIFAR, and ImageNet, which have standard splits, but it does not explicitly specify the training, validation, or test split percentages or sample counts used for its experiments.
Hardware Specification Yes All experiments are conducted in the workstation with four NVIDIA RTX A6000 GPUs, and take about one day to run.
Software Dependencies No The paper does not provide specific software names with version numbers (e.g., Python, PyTorch, or other libraries with their versions) that were used in the experiments.
Experiment Setup Yes For SZOHT, we set k = 10, s2 = 10, q = 10, and (µ, η) = (0.015, 0.015) for port4, and (µ, η) = (0.1, 1) for port5 (µ and η are both obtained by grid search over the interval [10 3, 103]). We set the hyperparameters of SZOHT as follows: MNIST: k = 20, s2 = 100, q = 100, µ = 0.3, η = 1; CIFAR: k = 60, s2 = 100, q = 1000, µ = 1e 3, η = 0.01; Image Net: k = 100000, s2 = 1000, q = 100, µ = 0.01, η = 0.015.