[논문 리뷰] NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture\n Search

Xuanyi Dong, Yi Yang|arXiv (Cornell University)|2020. 01. 02.

Advanced Neural Network Applications인용 수 290

한 줄 요약

이 논문은 NAS-Bench-201을 소개합니다. 고정된 셀 기반 탐색 공간 벤치마크로서 15,625개의 아키텍처를 세 가지 데이터셋에서 평가하고, 공정한 NAS 알고리즘 비교와 학습된 모델 매개변수의 재사용을 가능하게 하는 풍부한 훈련 로그 및 진단 데이터를 제공합니다.

ABSTRACT

Neural architecture search (NAS) has achieved breakthrough success in a great\nnumber of applications in the past few years. It could be time to take a step\nback and analyze the good and bad aspects in the field of NAS. A variety of\nalgorithms search architectures under different search space. These searched\narchitectures are trained using different setups, e.g., hyper-parameters, data\naugmentation, regularization. This raises a comparability problem when\ncomparing the performance of various NAS algorithms. NAS-Bench-101 has shown\nsuccess to alleviate this problem. In this work, we propose an extension to\nNAS-Bench-101: NAS-Bench-201 with a different search space, results on multiple\ndatasets, and more diagnostic information. NAS-Bench-201 has a fixed search\nspace and provides a unified benchmark for almost any up-to-date NAS\nalgorithms. The design of our search space is inspired from the one used in the\nmost popular cell-based searching algorithms, where a cell is represented as a\nDAG. Each edge here is associated with an operation selected from a predefined\noperation set. For it to be applicable for all NAS algorithms, the search space\ndefined in NAS-Bench-201 includes all possible architectures generated by 4\nnodes and 5 associated operation options, which results in 15,625 candidates in\ntotal. The training log and the performance for each architecture candidate are\nprovided for three datasets. This allows researchers to avoid unnecessary\nrepetitive training for selected candidate and focus solely on the search\nalgorithm itself. The training time saved for every candidate also largely\nimproves the efficiency of many methods. We provide additional diagnostic\ninformation such as fine-grained loss and accuracy, which can give inspirations\nto new designs of NAS algorithms. In further support, we have analyzed it from\nmany aspects and benchmarked 10 recent NAS algorithms.\n

연구 동기 및 목표

• Provide a fixed, algorithm-agnostic cell-based search space for reproducible NAS evaluations.
• Offer training logs, accuracy, and loss data across multiple datasets for each architecture to avoid redundant training.
• Deliver additional diagnostic information (parameters, FLOPs, latency, fine-grained training dynamics) to inform NAS design.
• Enable benchmarking and fair comparison of diverse NAS algorithms on standardized datasets.

제안 방법

• Define a fixed 4-node densely connected DAG cell with 5 operation options per edge (zeroize, skip, 1x1 conv, 3x3 conv, 3x3 avg pool).
• Assemble cells into a standard macro skeleton with three stages and a residual downsampling block comparable to modern cell-based NAS methods.
• Train and evaluate every architecture candidate (15,625 total) on CIFAR-10, CIFAR-100, and ImageNet-16-120 using a unified training protocol and epochs.
• Provide complete training logs (loss/accuracy per epoch), parameter counts, FLOPs, and latency for each architecture & run.
• Benchmark 10 NAS algorithms (RL, ES, differentiable, HPO) on NAS-Bench-201 to assess speedups and transferability.]
• research_questions: ["How does a fixed, comprehensive NAS cell space enable fairer comparison across NAS algorithms?","What is the performance distribution and transferability of architectures across CIFAR-10, CIFAR-100, and ImageNet-16-120 within NAS-Bench-201?","Can NAS algorithms achieve substantial speedups using the benchmark’s precomputed results without retraining architectures?","What diagnostic information can inform the design of more robust and efficient NAS methods?"]
• key_findings: ["The NAS-Bench-201 space contains 15,625 candidate cells and supports evaluations across three datasets.","- A wide range of NAS methods can be benchmarked in a unified, reproducible framework with consistent training setups.","DARTS variants and parameter-sharing approaches show sensitivity to BN handling and training hyper-parameters within this space.","Non-parameter-sharing NAS methods (REA, RS, REINFORCE, BOHB) can achieve strong rankings with fine-grained training signals and shorter runs.","The benchmark reveals generally consistent architecture rankings across datasets and highlights transferability challenges when directly porting architectures between datasets."]
• table_headers: [],
• table_rows: []

실험 결과

연구 질문

• RQ1 How does a fixed, comprehensive NAS cell space enable fairer comparison across NAS algorithms?
• RQ2 What is the performance distribution and transferability of architectures across CIFAR-10, CIFAR-100, and ImageNet-16-120 within NAS-Bench-201?
• RQ3 Can NAS algorithms achieve substantial speedups using the benchmark’s precomputed results without retraining architectures?
• RQ4 What diagnostic information can inform the design of more robust and efficient NAS methods?

주요 결과

• The NAS-Bench-201 space contains 15,625 candidate cells and supports evaluations across three datasets.
• - A wide range of NAS methods can be benchmarked in a unified, reproducible framework with consistent training setups.
• DARTS variants and parameter-sharing approaches show sensitivity to BN handling and training hyper-parameters within this space.
• Non-parameter-sharing NAS methods (REA, RS, REINFORCE, BOHB) can achieve strong rankings with fine-grained training signals and shorter runs.
• The benchmark reveals generally consistent architecture rankings across datasets and highlights transferability challenges when directly porting architectures between datasets.