Understanding deep neural network by filter sensitive area generation network

Yang Qian; Hong Qiao; Jing Xu

doi:10.1007/978-3-030-04167-0_18

Understanding deep neural network by filter sensitive area generation network

Yang Qian, Hong Qiao, Jing Xu

Research output: Chapters, Conference Papers, Creative and Literary Works › RGC 32 - Refereed conference paper (with host publication) › peer-review

1 Citation (Scopus)

Abstract

Deep convolutional networks have recently gained much attention because of their impressive performance on some visual tasks. However, it is still not clear why they achieve such great success. In this paper, a novel approach called Filter Sensitive Area Generation Network (FSAGN), has been proposed to interpret what the convolutional filters have learnt after training CNNs. Given any trained CNN model, the proposed method aims to figure out which object part each filter represents in a high conv-layer, through appropriate input image mask which filters out unrelated area. In order to obtain such a mask, a mask generation network is designed and the corresponding loss function is defined to evaluate the changes of feature maps before and after mask operation. Experiments on multiple datasets and networks show that FSAGN clarifies the knowledge representations of each filter and how small disturbance on specific object parts affects the performance of CNNs. © 2018, Springer Nature Switzerland AG.

Original language	English
Title of host publication	Neural Information Processing - 25th International Conference, ICONIP 2018, Proceedings
Publisher	Springer Verlag
Pages	192-203
Volume	11301 LNCS
ISBN (Print)	9783030041663
DOIs	https://doi.org/10.1007/978-3-030-04167-0_18
Publication status	Published - 2018
Externally published	Yes
Event	25th International Conference on Neural Information Processing, ICONIP 2018 - Siem Reap, Cambodia Duration: 13 Dec 2018 → 16 Dec 2018

Publication series

Name	Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume	11301 LNCS
ISSN (Print)	0302-9743
ISSN (Electronic)	1611-3349

Conference

Conference	25th International Conference on Neural Information Processing, ICONIP 2018
Place	Cambodia
City	Siem Reap
Period	13/12/18 → 16/12/18

Bibliographical note

Publication details (e.g. title, author(s), publication statuses and dates) are captured on an “AS IS” and “AS AVAILABLE” basis at the time of record harvesting from the data source. Suggestions for further amendments or supplementary information can be sent to <a href="mailto:[email protected]">[email protected]</a>.

Funding

This work was supported in part by the National Key Research and Development Program of China (2017YFB1300203), in part by the National Natural Science Foundation of China under Grant 91648205.

Research Keywords

Convolutional neural network
Interpretability
Knowledge representations

Access Output

10.1007/978-3-030-04167-0_18

Other Access Options

Check CityUHK Library

Permanent Link

Right click to copy link

Cite this

Qian, Y., Qiao, H., & Xu, J. (2018). Understanding deep neural network by filter sensitive area generation network. In Neural Information Processing - 25th International Conference, ICONIP 2018, Proceedings (Vol. 11301 LNCS, pp. 192-203). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 11301 LNCS). Springer Verlag. https://doi.org/10.1007/978-3-030-04167-0_18

Qian, Yang ; Qiao, Hong ; Xu, Jing. / Understanding deep neural network by filter sensitive area generation network. Neural Information Processing - 25th International Conference, ICONIP 2018, Proceedings. Vol. 11301 LNCS Springer Verlag, 2018. pp. 192-203 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)).

@inproceedings{c503500621254ccb88c7e90e84ca6588,

title = "Understanding deep neural network by filter sensitive area generation network",

abstract = "Deep convolutional networks have recently gained much attention because of their impressive performance on some visual tasks. However, it is still not clear why they achieve such great success. In this paper, a novel approach called Filter Sensitive Area Generation Network (FSAGN), has been proposed to interpret what the convolutional filters have learnt after training CNNs. Given any trained CNN model, the proposed method aims to figure out which object part each filter represents in a high conv-layer, through appropriate input image mask which filters out unrelated area. In order to obtain such a mask, a mask generation network is designed and the corresponding loss function is defined to evaluate the changes of feature maps before and after mask operation. Experiments on multiple datasets and networks show that FSAGN clarifies the knowledge representations of each filter and how small disturbance on specific object parts affects the performance of CNNs. {\textcopyright} 2018, Springer Nature Switzerland AG.",

keywords = "Convolutional neural network, Interpretability, Knowledge representations",

author = "Yang Qian and Hong Qiao and Jing Xu",

note = "Publication details (e.g. title, author(s), publication statuses and dates) are captured on an “AS IS” and “AS AVAILABLE” basis at the time of record harvesting from the data source. Suggestions for further amendments or supplementary information can be sent to <a href={"}mailto:[email protected]{"}>[email protected]</a>.; 25th International Conference on Neural Information Processing, ICONIP 2018 ; Conference date: 13-12-2018 Through 16-12-2018",

year = "2018",

doi = "10.1007/978-3-030-04167-0_18",

language = "English",

isbn = "9783030041663",

volume = "11301 LNCS",

series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",

publisher = "Springer Verlag",

pages = "192--203",

booktitle = "Neural Information Processing - 25th International Conference, ICONIP 2018, Proceedings",

address = "Germany",

}

Qian, Y, Qiao, H & Xu, J 2018, Understanding deep neural network by filter sensitive area generation network. in Neural Information Processing - 25th International Conference, ICONIP 2018, Proceedings. vol. 11301 LNCS, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 11301 LNCS, Springer Verlag, pp. 192-203, 25th International Conference on Neural Information Processing, ICONIP 2018, Siem Reap, Cambodia, 13/12/18. https://doi.org/10.1007/978-3-030-04167-0_18

Understanding deep neural network by filter sensitive area generation network. / Qian, Yang; Qiao, Hong; Xu, Jing.
Neural Information Processing - 25th International Conference, ICONIP 2018, Proceedings. Vol. 11301 LNCS Springer Verlag, 2018. p. 192-203 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 11301 LNCS).

Research output: Chapters, Conference Papers, Creative and Literary Works › RGC 32 - Refereed conference paper (with host publication) › peer-review

TY - GEN

T1 - Understanding deep neural network by filter sensitive area generation network

AU - Qian, Yang

AU - Qiao, Hong

AU - Xu, Jing

N1 - Publication details (e.g. title, author(s), publication statuses and dates) are captured on an “AS IS” and “AS AVAILABLE” basis at the time of record harvesting from the data source. Suggestions for further amendments or supplementary information can be sent to <a href="mailto:[email protected]">[email protected]</a>.

PY - 2018

Y1 - 2018

N2 - Deep convolutional networks have recently gained much attention because of their impressive performance on some visual tasks. However, it is still not clear why they achieve such great success. In this paper, a novel approach called Filter Sensitive Area Generation Network (FSAGN), has been proposed to interpret what the convolutional filters have learnt after training CNNs. Given any trained CNN model, the proposed method aims to figure out which object part each filter represents in a high conv-layer, through appropriate input image mask which filters out unrelated area. In order to obtain such a mask, a mask generation network is designed and the corresponding loss function is defined to evaluate the changes of feature maps before and after mask operation. Experiments on multiple datasets and networks show that FSAGN clarifies the knowledge representations of each filter and how small disturbance on specific object parts affects the performance of CNNs. © 2018, Springer Nature Switzerland AG.

AB - Deep convolutional networks have recently gained much attention because of their impressive performance on some visual tasks. However, it is still not clear why they achieve such great success. In this paper, a novel approach called Filter Sensitive Area Generation Network (FSAGN), has been proposed to interpret what the convolutional filters have learnt after training CNNs. Given any trained CNN model, the proposed method aims to figure out which object part each filter represents in a high conv-layer, through appropriate input image mask which filters out unrelated area. In order to obtain such a mask, a mask generation network is designed and the corresponding loss function is defined to evaluate the changes of feature maps before and after mask operation. Experiments on multiple datasets and networks show that FSAGN clarifies the knowledge representations of each filter and how small disturbance on specific object parts affects the performance of CNNs. © 2018, Springer Nature Switzerland AG.

KW - Convolutional neural network

KW - Interpretability

KW - Knowledge representations

UR - http://www.scopus.com/inward/record.url?scp=85059075274&partnerID=8YFLogxK

UR - https://www.scopus.com/record/pubmetrics.uri?eid=2-s2.0-85059075274&origin=recordpage

U2 - 10.1007/978-3-030-04167-0_18

DO - 10.1007/978-3-030-04167-0_18

M3 - RGC 32 - Refereed conference paper (with host publication)

SN - 9783030041663

VL - 11301 LNCS

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 192

EP - 203

BT - Neural Information Processing - 25th International Conference, ICONIP 2018, Proceedings

PB - Springer Verlag

T2 - 25th International Conference on Neural Information Processing, ICONIP 2018

Y2 - 13 December 2018 through 16 December 2018

ER -

Qian Y, Qiao H, Xu J. Understanding deep neural network by filter sensitive area generation network. In Neural Information Processing - 25th International Conference, ICONIP 2018, Proceedings. Vol. 11301 LNCS. Springer Verlag. 2018. p. 192-203. (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)). doi: 10.1007/978-3-030-04167-0_18

Understanding deep neural network by filter sensitive area generation network

Abstract

Publication series

Conference

Bibliographical note

Funding

Research Keywords

Access Output

Other Access Options

Permanent Link

Other Files and Links

Fingerprint

Cite this