Group Differentiable Architecture Search
| Autor: | Jinhua Xu, Chaoyuan Shen |
|---|---|
| Jazyk: | angličtina |
| Rok vydání: | 2021 |
| Předmět: |
Theoretical computer science
shallow architecture General Computer Science Computer science media_common.quotation_subject 010501 environmental sciences Space (commercial competition) Overfitting 01 natural sciences 0502 economics and business General Materials Science Simplicity Differentiable function 050207 economics Electrical and Electronic Engineering Architecture 0105 earth and related environmental sciences media_common Group (mathematics) 05 social sciences General Engineering Unfair competition skip-connect aggregation TK1-9971 Enhanced Data Rates for GSM Evolution Electrical engineering. Electronics. Nuclear engineering Differentiable architecture search |
| Zdroj: | IEEE Access, Vol 9, Pp 76585-76591 (2021) |
| ISSN: | 2169-3536 |
| Popis: | Differentiable architecture search (DARTS) has received great attention due to its simplicity and efficiency. However, there are two annoying problems. One is that searched architecture of normal cell tends to be shallow. The other is skip-connect aggregation caused by the unfair competition between operations. We find that fewer operations per edge is helpful to search for deeper architectures, so we divide the operations into groups and train these groups in turn. To explore competitiveness among all the operations in the search space, candidate operations will be regrouped in each epoch. In addition, the random grouping prevents the overfitting of the super network, and consequently avoids the skip-connect aggregation. We named this method GroupDARTS and evaluated these searched architectures, achieving a state-of-the-art result of 97.68% on CIFAR10 and a top-1 accuray of 75.5% on ImageNet. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|