PaperReading20200220

CanChen ggchen@mail.ustc.edu.cn

---

Recently I am occupied with something else so today I am going to share only one paper.

 

Rule extraction from RNN

• Motivation:RNN is good at handling sequence modeling problems and researchers try to extract DFA from RNN hidden states. In this paper, the author wants to find out factors that influence the DFA extracted from RNN.
• Method: The author first trains a second-order RNN and then uses k-means to cluster the hidden states. At last, the author bulids the state transition and it is simplied to DFA.
• Contribution: This paper shows us that sometimes the extracted DFA can be more accurate than the original RNN model. What's more, DFA often requires less effort for training. To conclude, this work belongs to explaining blackbox models.