Please wait a minute...
FITEE
Front. Inform. Technol. Electron. Eng.  2018, Vol. 19 Issue (1): 27-39    
    
Visual interpretability for deep learning: a survey
Quan-shi ZHANG, Song-chun ZHU
University of California, Los Angeles, California 90095, USA
Visual interpretability for deep learning: a survey
Quan-shi ZHANG, Song-chun ZHU
University of California, Los Angeles, California 90095, USA
 全文: PDF 
摘要： This paper reviews recent studies in understanding neural-network representations and learning neu-
ral  networks with interpretable/disentangled  middle-layer representations.   Although deep neural networks have
exhibited superior performance in various tasks, interpretability is always Achilles’ heel of deep neural networks.
At present, deep neural networks obtain high discrimination power at the cost of a low interpretability of their
black-box representations. We believe that high model interpretability may help people break several bottlenecks of
deep learning, e.g., learning from a few annotations, learning via human–computer communications at the semantic
level,  and semantically debugging network representations.   We focus on convolutional neural networks (CNNs),
and revisit the visualization of CNN representations, methods of diagnosing representations of pre-trained CNNs,
approaches for disentangling pre-trained CNN representations, learning of CNNs with disentangled representations,
and middle-to-end learning based on model interpretability.  Finally, we discuss prospective trends in explainable
artificial intelligence.
关键词： Artificial intelligence;  Deep learning;  Interpretable model    
Abstract: This paper reviews recent studies in understanding neural-network representations and learning neu-
ral  networks with interpretable/disentangled  middle-layer representations.   Although deep neural networks have
exhibited superior performance in various tasks, interpretability is always Achilles’ heel of deep neural networks.
At present, deep neural networks obtain high discrimination power at the cost of a low interpretability of their
black-box representations. We believe that high model interpretability may help people break several bottlenecks of
deep learning, e.g., learning from a few annotations, learning via human–computer communications at the semantic
level,  and semantically debugging network representations.   We focus on convolutional neural networks (CNNs),
and revisit the visualization of CNN representations, methods of diagnosing representations of pre-trained CNNs,
approaches for disentangling pre-trained CNN representations, learning of CNNs with disentangled representations,
and middle-to-end learning based on model interpretability.  Finally, we discuss prospective trends in explainable
artificial intelligence.
Key words: Artificial intelligence    Deep learning    Interpretable model
收稿日期: 2017-12-02 出版日期: 2019-06-06
服务　　
把本文推荐给朋友
加入引用管理器
E-mail Alert
RSS
作者相关文章　　
Quan-shi ZHANG
Song-chun ZHU

引用本文:

Quan-shi ZHANG, Song-chun ZHU. Visual interpretability for deep learning: a survey. Front. Inform. Technol. Electron. Eng., 2018, 19(1): 27-39.

链接本文:

http://www.zjujournals.com/xueshu/fitee/CN/    或    http://www.zjujournals.com/xueshu/fitee/CN/Y2018/V19/I1/27

[1] Heung-yeung SHUM, Xiao-dong HE, Di LI . From Eliza to XiaoIce: challenges and opportunities with social chatbots[J]. Front. Inform. Technol. Electron. Eng., 2018, 19(1): 10-26.
[2] Bin YU, Karl KUMBIER. Artificial intelligence and statistics[J]. Front. Inform. Technol. Electron. Eng., 2018, 19(1): 6-9.
[3] Partha Pratim ROY , Guoqiang ZHONG , Mohamed CHERIET. Tandem hidden Markov models using deep belief networks for offline handwriting recognition[J]. Front. Inform. Technol. Electron. Eng., 2017, 18(7): 978-988.