PaperReading20200226

CanChen ggchen@mail.ustc.edu.cn

---

 

To share or not share

• Motivation: With the publiaction of NAS101, the author wants to study how weight sharing work in different search spaces and the correlation between the supernet and results returned by NAS101.
• Method: The paper did extensive experiments in weight sharing methods with different search spaces and showed that search space plays an important role in the correlation.
• Contribution: The work made good use of NAS101 but I did not see any novelty in this work.

 

NAS EVALUATION IS FRUSTRATINGLY HARD

• Motivation: NAS needs some criteria.
• Method: This paper did a wide range of experiments and showed that cell-based searched models tends to have similar accuracies.
  Also, the selection of seeds and macro structures are important. The operations, instead, have less impact on the performance.
• Contribution: This paper is a ILCR paper, probably because this paper analysed some important issues in NAS evaluation.