Generative adversarial network for visualizing convolutional network

Masayuki Kobayashi, Masanori Suganuma, Tomoharu Nagao

Research output: Chapter in Book/Report/Conference proceedingConference contribution

1 Citation (Scopus)

Abstract

The convolutional neural network (CNN) is one of the most powerful models that has been achieving state-of-the art performance on a variety of computer vision tasks. However, their models are often considered as black-boxes and their lack of interpretability are considered to be a major problem. For example, in applications where the high interpretability is important, getting the predictions is not enough and we also need understanding of why these predictions are made. In that sense, understanding how their models work can build trust with users and make them more human-oriented. In this paper, we introduce a new visualization framework based on generative adversarial network (GAN) to give an insight into the neuron activities and what the CNN has learned. Our method is very simple, yet can produce recognizable and interpretable visualizations. Applied our method to AlexNet, these visualizations help to understand how the CNN works internally.

Original languageEnglish
Title of host publication2017 IEEE 10th International Workshop on Computational Intelligence and Applications, IWCIA 2017 - Proceedings
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages153-158
Number of pages6
ISBN (Electronic)9781538604694
DOIs
Publication statusPublished - 2017 Dec 13
Externally publishedYes
Event10th IEEE International Workshop on Computational Intelligence and Applications, IWCIA 2017 - Hiroshima, Japan
Duration: 2017 Nov 112017 Nov 12

Publication series

Name2017 IEEE 10th International Workshop on Computational Intelligence and Applications, IWCIA 2017 - Proceedings
Volume2017-December

Conference

Conference10th IEEE International Workshop on Computational Intelligence and Applications, IWCIA 2017
CountryJapan
CityHiroshima
Period17/11/1117/11/12

Keywords

  • convolutional neural network
  • generative adversarial network
  • viziualozation

ASJC Scopus subject areas

  • Artificial Intelligence
  • Hardware and Architecture
  • Software
  • Control and Optimization

Fingerprint Dive into the research topics of 'Generative adversarial network for visualizing convolutional network'. Together they form a unique fingerprint.

Cite this