human automl
search space limited exhaustive
proxy-based
step#1: define search space: conv cell
normal cell \ reduction cell
step#2:learn the optimal cell architecture for proxy task : CIFAR-10
step#3 : transfer the learned optimal architecture to target task: ImageNet
weight sharing-based approach
conclusion:
automl极大提高了生产效率:allows non-experts to solve ml problems
automl在很多AI任务上取得了比手动设计更好的性能(exceeds the world’s best ml experts ar many different tasks)
automl是否能构造出经典的模型结构?
由于automl先要define一个search space design,而这个define一般是借鉴经典模型,还是建立在已有经典模型上;