PaperReading20200222

CanChen ggchen@mail.ustc.edu.cn

---

 

VS-GAE

• Motivation: With the publication of NAS101, researchers can do NAS research easily. The cell design problem is in fact a graph learning problem.Thus the paper proposes a auto-encoder based on graphs, uses it to learn the embedding space and generate new cells, namely, graphs.
• Method: The method is quite similar to the traditional auto-encoder method except using GNN as the encoder and the decoder. Plus, the graph generation process is a sequential process.
• Contribution: The paper is not very novel.

 

SimCLR

• Motivation: Current self-supervised tasks need some hand-designed pretext tasks and this paper proposes some data augmentations.By trying to find the invariant nature of these images, the model learn useful weights.
• Method: This paper proposes some data augmentation methods and adds a non-linear projection layer. The loss function is designed by sampling positive and negative pairs in the same batch. The positive pairs is an example image and its corresponding image under some purturbation.
• Contribution: This paper is helpful for engineering, showing us bigger batchsize and larger model lead to better results for SimCLR.