One-Shot Neural Architecture Search: Maximising Diversity to Overcome Catastrophic Forgetting
文献类型：期刊论文
| 作者 | Zhang, Miao4,5,6; Li, Huiqi6; Pan, Shirui5; Chang, Xiaojun3,5; Zhou, Chuan2 ; Ge, Zongyuan1; Su, Steven4
|
| 刊名 | IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
 |
| 出版日期 | 2021-09-01 |
| 卷号 | 43期号:9页码:2921-2935 |
| 关键词 | Computer architecture Training Optimization Neural networks Search methods Australia Germanium AutoML neural architecture search continual learning catastrophic forgetting novelty search |
| ISSN号 | 0162-8828 |
| DOI | 10.1109/TPAMI.2020.3035351 |
| 英文摘要 | One-shot neural architecture search (NAS) has recently become mainstream in the NAS community because it significantly improves computational efficiency through weight sharing. However, the supernet training paradigm in one-shot NAS introduces catastrophic forgetting, where each step of the training can deteriorate the performance of other architectures that contain partially-shared weights with current architecture. To overcome this problem of catastrophic forgetting, we formulate supernet training for one-shot NAS as a constrained continual learning optimization problem such that learning the current architecture does not degrade the validation accuracy of previous architectures. The key to solving this constrained optimization problem is a novelty search based architecture selection (NSAS) loss function that regularizes the supernet training by using a greedy novelty search method to find the most representative subset. We applied the NSAS loss function to two one-shot NAS baselines and extensively tested them on both a common search space and a NAS benchmark dataset. We further derive three variants based on the NSAS loss function, the NSAS with depth constrain (NSAS-C) to improve the transferability, and NSAS-G and NSAS-LG to handle the situation with a limited number of constraints. The experiments on the common NAS search space demonstrate that NSAS and it variants improve the predictive ability of supernet training in one-shot NAS with remarkable and efficient performance on the CIFAR-10, CIFAR-100, and ImageNet datasets. The results with the NAS benchmark dataset also confirm the significant improvements these one-shot NAS baselines can make. |
| 资助项目 | NSFC[61702415] ; NSFC[61972315] ; Australian Research Council (ARC) under a Discovery Early Career Researcher Award (DECRA)[DE190100626] ; Air Force Research Laboratory, DARPA[FA8750-19-20501] ; Youth Innovation Promotion Association CAS[2017210] |
| WOS研究方向 | Computer Science ; Engineering |
| 语种 | 英语 |
| WOS记录号 | WOS:000681124300008 |
| 出版者 | IEEE COMPUTER SOC |
| 源URL | [http://ir.amss.ac.cn/handle/2S8OKBNM/59032]  |
| 专题 | 应用数学研究所 |
| 通讯作者 | Li, Huiqi; Pan, Shirui |
| 作者单位 | 1.Monash Univ, Monash E Res Ctr, Clayton, Vic 3800, Australia 2.Chinese Acad Sci, Acad Math & Syst Sci, Beijing 100081, Peoples R China 3.King Abdulaziz Univ, Fac Comp & Informat Technol, Jeddah 21589, Saudi Arabia 4.Univ Technol Sydney, Fac Engn & Informat Technol, Ultimo, NSW 2007, Australia 5.Monash Univ, Fac Informat Technol, Clayton, Vic 3800, Australia 6.Beijing Inst Technol, Sch Informat & Elect, Beijing 100081, Peoples R China |
| 推荐引用方式 GB/T 7714 | Zhang, Miao,Li, Huiqi,Pan, Shirui,et al. One-Shot Neural Architecture Search: Maximising Diversity to Overcome Catastrophic Forgetting[J]. IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE,2021,43(9):2921-2935. |
| APA | Zhang, Miao.,Li, Huiqi.,Pan, Shirui.,Chang, Xiaojun.,Zhou, Chuan.,...&Su, Steven.(2021).One-Shot Neural Architecture Search: Maximising Diversity to Overcome Catastrophic Forgetting.IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE,43(9),2921-2935. |
| MLA | Zhang, Miao,et al."One-Shot Neural Architecture Search: Maximising Diversity to Overcome Catastrophic Forgetting".IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE 43.9(2021):2921-2935. |
入库方式： OAI收割
来源：数学与系统科学研究院
浏览0
下载0
收藏0
其他版本
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。