PaperReading20200224

CanChen ggchen@mail.ustc.edu.cn

---

 

BANANAS

• Motivation: This paper proposes a network performance predictor to ease the computing pressure in the bayesian optimization process for network architecture search.
• Method: The paper first uses one-hot path encoding method to represent a network architecure and then trains a accuracy regression model. After that, the author uses the model in the bayesian optimization process and updates it every iteration.
• Contribution: One-hot path encoding may be important.

 

Best Practices

• Motivation: NAS has been a hot topic these days but many papers reporting NAS results have some problems.
• Method: Releasing code in time and making every training detail available are very important.
• Contribution: Give machine learning researchers a standard guideline for their research.