A rule extraction study on a neural network trained by deep learning

Bologna, Guido (School of Engineering, Architecture and Landscape (hepia), HES-SO // University of Applied Sciences Western Switzerland) ; Hayashi, Yoichi (Department of Computer Science, Meiji University, Kawasaki, Japan)

Rule extraction from neural networks is a fervent research topic. In the last 20 years many authors presented a number of techniques showing how to extract symbolic rules from Multi Layer Perceptrons (MLPs). Nevertheless, very few were related to ensembles of neural networks and even less for networks trained by deep learning. In this work the Discretized Multi Layer Perceptron (DIMLP) was trained by deep learning, then symbolic rules were extracted in an easier way with respect to standard MLPs. We compared the accuracy of deep trained DIMLPs and DIMLP ensembles on a subset of the MNIST dataset. The former networks were more accurate than the latter. Moreover, the complexity of the rules extracted from deep trained DIMLPs was similar to that obtained by boosted ensembles of DIMLPs. Finally, we examined the generated rules with respect to the centroids of the covered samples. Qualitatively, no clear difference in the strategy of classification emerged between deep trained DIMLPs and the ensembles.


Keywords:
Conference Type:
full paper
Faculty:
Ingénierie et Architecture
School:
HEPIA - Genève
Institute:
inIT - Institut d'Ingénierie Informatique et des Télécommunications
Publisher:
Vancouver, BC, Canada, 24-29 July 2016
Date:
2016-07
Vancouver, BC, Canada
24-29 July 2016
Pagination:
8 p.
Published in:
Proceedings of 2016 International Joint Conference on Neural Networks (IJCNN), 24-29 July 2016, Vancouver, BC, Canada
Numeration (vol. no.):
pp. 668-675
DOI:
ISBN:
978-1-5090-0620-5
Appears in Collection:

Note: The status of this file is: restricted


 Record created 2020-07-10, last modified 2020-07-14

Fulltext:
Download fulltext
PDF

Rate this document:

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