A Novel Training Protocol for Performance Predictors of Evolutionary Neural Architecture Search Algorithms

Evolutionary neural architecture search (ENAS) can automatically design the architectures of deep networks (DNNs) using evolutionary computation algorithms. However, most ENAS algorithms require an intensive computational resource, which is not necessarily available to users interested. Performance predictors are a type regression models assist accomplish search, while without exerting much resource. Despite various performance have been designed, they employ same training protocol build models: 1) sampling set DNNs with as dataset; 2) model mean square error criterion; and 3) predicting newly generated during ENAS. In this article, we point out that three steps constituting well thought-out through intuitive illustrative examples. Furthermore, propose new address these issues, consisting designing pairwise ranking indicator construct target, proposing use logistic fit samples, developing differential method instances. To verify effectiveness proposed protocol, four widely used in field machine learning chosen perform comparisons on two benchmark datasets. The experimental results all demonstrate significantly improve prediction accuracy against traditional protocols.