Optimal Transport Kernels for Sequential and Parallel Neural Architecture Search

Authors: Vu Nguyen, Tam Le, Makoto Yamada, Michael A. Osborne

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically demonstrate that our TW-based approaches outperform other baselines in both sequential and parallel NAS.
Researcher Affiliation Collaboration 1Amazon Adelaide (work done prior to joining Amazon) 2RIKEN AIP 3Kyoto University 4University of Oxford.
Pseudocode Yes Algorithm 1 Sequential and Parallel NAS using Gaussian process with tree-Wasserstein kernel
Open Source Code Yes We release the Python code for our experiments at https://github.com/ntienvu/TW_NAS.
Open Datasets Yes We utilize the popular NAS tabular datasets of Nasbench101 (NB101) (Ying et al., 2019) and Nasbench201 (NB201) (Dong & Yang, 2020) for evaluations.
Dataset Splits No The paper mentions allocating queries for NAS benchmarks but does not specify the train/validation/test splits of the underlying classification datasets (e.g., CIFAR-10) that these benchmarks utilize.
Hardware Specification No The paper acknowledges 'NVIDIA for sponsoring GPU hardware and Google Cloud Platform for sponsoring computing resources'. However, it does not specify exact GPU models, CPU models, or other detailed hardware specifications.
Software Dependencies No The paper mentions using the 'POT library (Flamary & Courty, 2017)' and the 'Auto ML library for TPE and BOHB3'. However, specific version numbers for these software dependencies are not provided.
Experiment Setup Yes All experimental results are averaged over 30 independent runs with different random seeds. We set the number of candidate architecture |Pt| = 100. We allocate a maximum budget of 500 queries for NB101 and 200 queries for NB201 including 10% of random selection at the beginning of BO.