IT Home August 20th MIUI official today announced that the Xiaomi AutoML team has created the FairNAS, MoGA, and SCARLET trilogy, surpassing MnasNet, MobileNetV3, and EfficientNet of Google's top teams.
The official said the August 19, millet AutoML team (authors: the beginning of Xiang Xiang, Zhang Bo, Li Jixiang, Li Qingyuan, Xu Ruijun) showcase the latest achievements SCARLET again, the same order of magnitude than Google EfficientNet model Brain chief scientist Quoc Le team proposed (ICML 2019).
The paper proposes an automated neural network search with telescopic function, which solves the problem of fairness of scalable supernet training in one-shot route through linear equivalent transformation. This release of the SCARLET series is directly benchmarked to EfficientNet, achieving 76.9% Top-1 accuracy in the ImageNet 1k classification task, currently SOTA with a <390 M FLOPs rating. At this point, the Xiaomi AutoML team created the FairNAS, MoGA, and SCARLET trilogy, surpassing MnasNet, MobileNetV3, and EfficientNet of Google's top teams.