Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Efficient Few-Shot Neural Architecture Search by Counting the Number of Nonlinear Functions
Authors: Youngmin Oh, Hyunju Lee, Bumsub Ham
AAAI 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on standard NAS benchmarks demonstrate the effectiveness of our approach. ... We perform experiments on standard NAS benchmarks for image classification: CIFAR10 (Krizhevsky, Hinton et al. 2009) and Image Net (Krizhevsky, Sutskever, and Hinton 2012). |
| Researcher Affiliation | Academia | 1Yonsei University 2Korea Institute of Science and Technology (KIST) EMAIL |
| Pseudocode | No | The paper describes the methods in prose and through mathematical formulations but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code https://cvlab.yonsei.ac.kr/projects/EFS-NAS |
| Open Datasets | Yes | We perform experiments on standard NAS benchmarks for image classification: CIFAR10 (Krizhevsky, Hinton et al. 2009) and Image Net (Krizhevsky, Sutskever, and Hinton 2012). ... We adopt NAS201 (Dong and Yang 2020) and Mobile Net (Cai, Zhu, and Han 2019; Sandler et al. 2018) search spaces for CIFAR10 and Image Net, respectively. |
| Dataset Splits | Yes | Following the standard protocol in (Liu, Simonyan, and Yang 2019; Xu et al. 2020; Guo et al. 2020), we split the training set of CIFAR10 in half and use each for training and validation, respectively. For Image Net, we sample 50K images from the training set to construct a new validation set, and use the original validation set for testing. |
| Hardware Specification | Yes | All experiments are performed with 8 NVIDIA A5000 GPUs. |
| Software Dependencies | No | The paper mentions optimizers (SGD, RMSProp) and a cosine annealing strategy but does not provide specific version numbers for any software libraries or dependencies. |
| Experiment Setup | Yes | Specifically, we train supernets for 200 epochs with a batch size of 1,024. We adopt a SGD optimizer with an initial learning rate of 0.12, a momentum of 0.9, and a weight decay of 4e-5. The learning rate is adjusted by a cosine annealing strategy without restart. The number of supernets K is set to 3 and 6 on NAS201 and Image Net, respectively. We set G to 2 for all experiments, unless otherwise specified. After applying the evolutionary search algorithm (Guo et al. 2020), we train the chosen architectures for 450 epochs with a batch size of 1,024. We use a RMSProp optimizer with an initial learning rate and a weight decay of 0.064 and 1e-5, respectively. The learning rate decays by a factor of 0.97 per 2.4 epochs. |