TabNAS: Rejection Sampling for Neural Architecture Search on Tabular Datasets

Authors: Chengrun Yang, Gabriel Bender, Hanxiao Liu, Pieter-Jan Kindermans, Madeleine Udell, Yifeng Lu, Quoc V Le, Da Huang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Results on several tabular datasets demonstrate the superiority of Tab NAS over previous reward-shaping methods: it finds better models that obey the constraints.
Researcher Affiliation Collaboration Chengrun Yang1, Gabriel Bender1, Hanxiao Liu1, Pieter-Jan Kindermans1, Madeleine Udell2, Yifeng Lu1, Quoc V. Le1, Da Huang1 {chengrun, gbender, hanxiaol, pikinder}@google.com, udell@stanford.edu, {yifenglu, qvl, dahua}@google.com 1 Google Research, Brain Team 2 Stanford University
Pseudocode Yes detailed pseudocode is provided as Algorithm 2 in Appendix B.
Open Source Code Yes Our implementation can be found at https://github.com/google-research/tabnas.
Open Datasets Yes The datasets are publicly available. We also provide pseudocode and full details of our hyperparameters to reproduce our results in Table A1 and A2.
Dataset Splits Yes To avoid overfitting, we split the labelled portion of a dataset into training and validation splits. Weight updates are carried out on the training split; RL updates are performed on the validation split.
Hardware Specification Yes We ran all experiments using Tensor Flow on a Cloud TPU v2 with 8 cores.
Software Dependencies No The paper mentions 'Tensor Flow' and briefly references 'Py Torch' but does not specify version numbers for these or any other software dependencies crucial for reproducibility.
Experiment Setup Yes More details of experiment setup and results in other search spaces can be found in Appendix C and D.