Generalizable Lightweight Proxy for Robust NAS against Diverse Perturbations

Authors: Hyeonjeong Ha, Minseon Kim, Sung Ju Hwang

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The experimental results demonstrate that our proxy can rapidly and efficiently search for neural architectures that are consistently robust against various perturbations on multiple benchmark datasets and diverse search spaces, largely outperforming existing clean zero-shot NAS and robust NAS with reduced search cost.
Researcher Affiliation Collaboration Hyeonjeong Ha1 , Minseon Kim1 , Sung Ju Hwang1,2 1Korea Advanced Institute of Science and Technology (KAIST), 2Deep Auto.ai {hyeonjeongha, minseonkim, sjhwang82}@kaist.ac.kr
Pseudocode No The paper describes the proposed method using mathematical equations but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https://github.com/HyeonjeongHa/CRoZe.
Open Datasets Yes Datasets. For the NAS-Bench-201 [12, 22] search space, we validate our proxy across different tasks (CIFAR-10, CIFAR-100, and Image Net16-120) and perturbations (FGSM [16], PGD [29], and 15 types of common corruptions [20]).
Dataset Splits Yes Datasets. For the NAS-Bench-201 [12, 22] search space, we validate our proxy across different tasks (CIFAR-10, CIFAR-100, and Image Net16-120) and perturbations (FGSM [16], PGD [29], and 15 types of common corruptions [20]). To measure Spearman s ρ between final accuracies and our proxy values, we use both clean NAS-Bench-201 [12] and robust NAS-Bench-201 [22], which include clean accuracies and robust accuracies, respectively. Finally, we search for the optimal architectures with our proxy in DARTS [26] search space and compare the final accuracies against previous NAS methods [31, 18, 6, 1] on CIFAR-10 and CIFAR-100.
Hardware Specification Yes All NAS methods are executed on a single NVIDIA 3090 RTX GPU.
Software Dependencies No The paper does not explicitly provide specific version numbers for software dependencies such as Python, PyTorch, or CUDA.
Experiment Setup Yes Standard Training: We train the neural architectures for 200 epochs under SGD optimizer with a learning rate of 0.1 and weight decay 1e-4, and use a batch size of 64 following [31].