Understanding Hidden Memories of Recurrent Neural Networks

被引:0
作者
Ming, Yao [1 ]
Cao, Shaozu [1 ]
Zhang, Ruixiang [1 ]
Li, Zhen [1 ]
Chen, Yuanzhe [1 ]
Song, Yangqiu [1 ]
Qu, Huamin [1 ]
机构
[1] Hong Kong Univ Sci & Technol, Hong Kong, Peoples R China
来源
2017 IEEE CONFERENCE ON VISUAL ANALYTICS SCIENCE AND TECHNOLOGY (VAST) | 2017年
关键词
recurrent neural networks; visual analytics; understanding neural model; co-clustering;
D O I
暂无
中图分类号
TM [电工技术]; TN [电子技术、通信技术];
学科分类号
0808 ; 0809 ;
摘要
Recurrent neural networks (RNNs) have been successfully applied to various natural language processing (NLP) tasks and achieved better results than conventional methods. However, the lack of understanding of the mechanisms behind their effectiveness limits further improvements on their architectures. In this paper, we present a visual analytics method for understanding and comparing RNN models for NLP tasks. We propose a technique to explain the function of individual hidden state units based on their expected response to input texts. We then co-cluster hidden state units and words based on the expected response and visualize co-clustering results as memory chips and word clouds to provide more structured knowledge on RNNs' hidden states. We also propose a glyph-based sequence visualization based on aggregate information to analyze the behavior of an RNN's hidden state at the sentence-level. The usability and effectiveness of our method are demonstrated through case studies and reviews from domain experts.
引用
收藏
页码:13 / 24
页数:12
相关论文
共 51 条
  • [1] Abadi Abadi M M, Tensorflow: Large-scale machine learning on heterogeneous distributed systems Tensorflow: Large-scale machine learning on heterogeneous distributed systems
  • [2] [Anonymous], 2015, NIPS
  • [3] [Anonymous], 2014, P 8 WORKSH SYNT SEM
  • [4] [Anonymous], 2015, ARXIV150602078
  • [5] [Anonymous], 2016, CoRR
  • [6] [Anonymous], COMPUTER GRAPHICS FO
  • [7] [Anonymous], P 5 INT C LEARN REPR
  • [8] [Anonymous], 1997, Neural Computation
  • [9] [Anonymous], 2016, ADAPT COMPUT MACH LE
  • [10] [Anonymous], CORR