PaperReading20200227

CanChen ggchen@mail.ustc.edu.cn

---

 

Neural Predictor for Neural Architecture Search

• Motivation: Current NAS algorithms require lots of computing resources.
• Method: This paper proposes a novel GCN-based network performance predictor and have done extensive experiments on ImageNet or NAS101.
• Contribution: This paper is not novel but in my opinion, network performance predictor is the future trend in accelerating NAS.

 

TAPAS

• Motivation: Current network performance methods do not take dataset difficulty into consideration.
• Method: The paper proposes a network performance predictor that can adjust to dataset difficulty. Another important thing is that this paper trains the network layer by layer.
• Contribution: What I learnt is that we can use NN to model a dataset fitting difficulty.