MathNAS: If Blocks Have a Role in Mathematical Architecture Design

Authors: Qinsi Wang, Jinghan Ke, Zhi Liang, Sihai Zhang

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

Reproducibility Variable Result LLM Response
Research Type Experimental The superiority of Math NAS is validated on multiple large-scale CV and NLP benchmark datasets. Notably on Image Net-1k, Math NAS achieves 82.5% top-1 accuracy, 1.2% and 0.96% higher than Swin-T and Le Vi T-256, respectively.
Researcher Affiliation Academia 1University of Science and Technology of China 2School of Life Sciences, University of Science and Technology of China 3Key Laboratory of Wireless-Optical Communications, Chinese Academy of Sciences 4School of Microelectronics, University of Science and Technology of China
Pseudocode Yes Algorithm 1 Math Neural Architecture Search
Open Source Code Yes Our code is available at https://github.com/wangqinsi1/Math NAS.
Open Datasets Yes The superiority of Math NAS is validated on multiple large-scale CV and NLP benchmark datasets. Notably on Image Net-1k, Math NAS achieves 82.5% top-1 accuracy...
Dataset Splits Yes For NAS-Bench-201 and Mobile Net V3, we adopt the training methodology employed in [16] and [22] to train the base net for 100 epochs. ... We sample 5 blocks per block node and count the accuracies of 3125 subnetworks on the Image Net validation set.
Hardware Specification Yes Raspberry Pi Intel Xeon CPU Nvidia TITAN Xp GPU
Software Dependencies No We employ the Gurobipy solver to address the ILP problem. No version number is specified for Gurobipy or any other software.
Experiment Setup No The settings of hyperparameters in the training are consistent with the original paper. No specific hyperparameter values (e.g., learning rate, batch size, optimizer settings) or explicit training configuration details are provided within this paper's main text.