Exploring internal representations of deep neural networks

Despraz, Jérémie (School of Management and Engineering Vaud, HES-SO // University of Applied Sciences Western Switzerland ; Computational Biology (CI4CB), SIB Swiss Institute of Bioinformatics, Lausanne, Switzerland) ; Gomez, Stéphane (School of Management and Engineering Vaud, HES-SO // University of Applied Sciences Western Switzerland ; Computational Intelligence for Computational Biology (CI4CB), SIB Swiss Institute of Bioinformatics, lausanne, Switzerland) ; Satizábal, Héctor F. (School of Management and Engineering Vaud, HES-SO // University of Applied Sciences Western Switzerland) ; Peña-Reyes, Carlos Andrés (School of Management and Engineering Vaud, HES-SO // University of Applied Sciences Western Switzerland ; Computational Intelligence for Computational Biology (CI4CB), SIB Swiss Institute of Bioinformatics, Lausanne, Switzerland)

This paper introduces a method for the generation of images that activate any target neuron or group of neurons of a trained convolutional neural network (CNN). These images are created in such a way that they contain attributes of natural images such as color patterns or textures. The main idea of the method is to pre-train a deep generative network on a dataset of natural images and then use this network to generate images for the target CNN. The analysis of the generated images allows for a better understanding of the CNN internal representations, the detection of otherwise unseen biases, or the creation of explanations through feature localization and description.


Keywords:
Faculty:
Ingénierie et Architecture
School:
HEIG-VD
Institute:
IICT - Institut des Technologies de l'Information et de la Communication
Publisher:
Cham, Springer
Date:
2019-05
Cham
Springer
Pagination:
20 p.
Published in:
Computational Intelligence (Studies in Computational Intelligence)
Author of the book:
Sabourin, Christophe
Merelo, Juan Julian
Madani, Kurosh
Warwick, Kevin
DOI:
ISSN:
1860-949X
ISBN:
978-3-030-16468-3
Appears in Collection:



 Record created 2019-11-26, last modified 2019-12-05


Rate this document:

Rate this document:
1
2
3
 
(Not yet reviewed)