| Presentation | 2019-12-06 Evaluation of the visualization techniques providing explanations for decisions of convolutional neural networks Mizuki Mori, Hiroki Tanaka, |
|---|---|
| PDF Download Page |  PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | Recent work has proposed a variety of techniques to visualize what a convolutional neural networks (CNN) utilizes to classify input images into particular classes. However, no established method can objectively evaluate which one provides better visual explanation. In this study, we present a method that directly evaluates the performance of the visualization method based only on changes in the output of the CNN, regardless of whether the visualized features are used by human for image classification. We devised a score measuring how much the output of the CNN's maximum response class decreases when the input is a masked image in which image region corresponding to the visualized features is masked, compared to when the input is an image without masking. By comparing the score when the mask is derived from the visualized features of the maximum response class with the score when the mask is derived from those for the remaining classes, the present method evaluates whether the visualization method produces features specifically relevant to individual classes. We applied the present method to several representative visualization algorithms. Our results suggest that there is a considerable difference in the explanation ability between them. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | Convolutional neural network / Visualization / Visual explanation |
| Paper # | MBE2019-59,NC2019-50 |
| Date of Issue | 2019-11-29 (MBE, NC) |
| Conference Information | |
| Committee | NC / MBE |
|---|---|
| Conference Date | 2019/12/6(1days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | Toyohashi Tech |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | |
| Chair | Hayaru Shouno(UEC) / Taishin Nomura(Osaka Univ.) |
| Vice Chair | Kazuyuki Samejima(Tamagawa Univ) / Takashi Watanabe(Tohoku Univ.) |
| Secretary | Kazuyuki Samejima(NAIST) / Takashi Watanabe(NTT) |
| Assistant | Takashi Shinozaki(NICT) / Ken Takiyama(TUAT) / Yasuyuki Suzuki(Osaka Univ.) / Akihiro Karashima(Tohoku Inst. of Tech.) |
| Paper Information | |
| Registration To | Technical Committee on Neurocomputing / Technical Committee on ME and Bio Cybernetics |
|---|---|
| Language | JPN |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | Evaluation of the visualization techniques providing explanations for decisions of convolutional neural networks |
| Sub Title (in English) | |
| Keyword(1) | Convolutional neural network |
| Keyword(2) | Visualization |
| Keyword(3) | Visual explanation |
| 1st Author's Name | Mizuki Mori |
| 1st Author's Affiliation | Kyoto-Sangyo University(Kyoto-Sangyo Univ) |
| 2nd Author's Name | Hiroki Tanaka |
| 2nd Author's Affiliation | Kyoto-Sangyo University(Kyoto-Sangyo Univ) |
| Date | 2019-12-06 |
| Paper # | MBE2019-59,NC2019-50 |
| Volume (vol) | vol.119 |
| Number (no) | MBE-327,NC-328 |
| Page | pp.pp.85-88(MBE), pp.85-88(NC), |
| #Pages | 4 |
| Date of Issue | 2019-11-29 (MBE, NC) |