Explaining nonlinear classification decisions with deep Taylor decomposition

Grégoire Montavon, Sebastian Lapuschkin, Alexander Binder, Wojciech Samek, Klaus Muller

Research output: Contribution to journalArticle

109 Citations (Scopus)

Abstract

Nonlinear methods such as Deep Neural Networks (DNNs) are the gold standard for various challenging machine learning problems such as image recognition. Although these methods perform impressively well, they have a significant disadvantage, the lack of transparency, limiting the interpretability of the solution and thus the scope of application in practice. Especially DNNs act as black boxes due to their multilayer nonlinear structure. In this paper we introduce a novel methodology for interpreting generic multilayer neural networks by decomposing the network classification decision into contributions of its input elements. Although our focus is on image classification, the method is applicable to a broad set of input data, learning tasks and network architectures. Our method called deep Taylor decomposition efficiently utilizes the structure of the network by backpropagating the explanations from the output to the input layer. We evaluate the proposed method empirically on the MNIST and ILSVRC data sets.

Original languageEnglish
Pages (from-to)211-222
Number of pages12
JournalPattern Recognition
Volume65
DOIs
Publication statusPublished - 2017 May 1

Fingerprint

Decomposition
Image recognition
Image classification
Multilayer neural networks
Network architecture
Transparency
Learning systems
Multilayers
Deep neural networks

Keywords

  • Deep neural networks
  • Heatmapping
  • Image recognition
  • Relevance propagation
  • Taylor decomposition

ASJC Scopus subject areas

  • Software
  • Signal Processing
  • Computer Vision and Pattern Recognition
  • Artificial Intelligence

Cite this

Explaining nonlinear classification decisions with deep Taylor decomposition. / Montavon, Grégoire; Lapuschkin, Sebastian; Binder, Alexander; Samek, Wojciech; Muller, Klaus.

In: Pattern Recognition, Vol. 65, 01.05.2017, p. 211-222.

Research output: Contribution to journalArticle

Montavon, G, Lapuschkin, S, Binder, A, Samek, W & Muller, K 2017, 'Explaining nonlinear classification decisions with deep Taylor decomposition', Pattern Recognition, vol. 65, pp. 211-222. https://doi.org/10.1016/j.patcog.2016.11.008
Montavon, Grégoire ; Lapuschkin, Sebastian ; Binder, Alexander ; Samek, Wojciech ; Muller, Klaus. / Explaining nonlinear classification decisions with deep Taylor decomposition. In: Pattern Recognition. 2017 ; Vol. 65. pp. 211-222.
@article{4a6ab8e3ed334802960e845aa1f87a32,
title = "Explaining nonlinear classification decisions with deep Taylor decomposition",
abstract = "Nonlinear methods such as Deep Neural Networks (DNNs) are the gold standard for various challenging machine learning problems such as image recognition. Although these methods perform impressively well, they have a significant disadvantage, the lack of transparency, limiting the interpretability of the solution and thus the scope of application in practice. Especially DNNs act as black boxes due to their multilayer nonlinear structure. In this paper we introduce a novel methodology for interpreting generic multilayer neural networks by decomposing the network classification decision into contributions of its input elements. Although our focus is on image classification, the method is applicable to a broad set of input data, learning tasks and network architectures. Our method called deep Taylor decomposition efficiently utilizes the structure of the network by backpropagating the explanations from the output to the input layer. We evaluate the proposed method empirically on the MNIST and ILSVRC data sets.",
keywords = "Deep neural networks, Heatmapping, Image recognition, Relevance propagation, Taylor decomposition",
author = "Gr{\'e}goire Montavon and Sebastian Lapuschkin and Alexander Binder and Wojciech Samek and Klaus Muller",
year = "2017",
month = "5",
day = "1",
doi = "10.1016/j.patcog.2016.11.008",
language = "English",
volume = "65",
pages = "211--222",
journal = "Pattern Recognition",
issn = "0031-3203",
publisher = "Elsevier Limited",

}

TY - JOUR

T1 - Explaining nonlinear classification decisions with deep Taylor decomposition

AU - Montavon, Grégoire

AU - Lapuschkin, Sebastian

AU - Binder, Alexander

AU - Samek, Wojciech

AU - Muller, Klaus

PY - 2017/5/1

Y1 - 2017/5/1

N2 - Nonlinear methods such as Deep Neural Networks (DNNs) are the gold standard for various challenging machine learning problems such as image recognition. Although these methods perform impressively well, they have a significant disadvantage, the lack of transparency, limiting the interpretability of the solution and thus the scope of application in practice. Especially DNNs act as black boxes due to their multilayer nonlinear structure. In this paper we introduce a novel methodology for interpreting generic multilayer neural networks by decomposing the network classification decision into contributions of its input elements. Although our focus is on image classification, the method is applicable to a broad set of input data, learning tasks and network architectures. Our method called deep Taylor decomposition efficiently utilizes the structure of the network by backpropagating the explanations from the output to the input layer. We evaluate the proposed method empirically on the MNIST and ILSVRC data sets.

AB - Nonlinear methods such as Deep Neural Networks (DNNs) are the gold standard for various challenging machine learning problems such as image recognition. Although these methods perform impressively well, they have a significant disadvantage, the lack of transparency, limiting the interpretability of the solution and thus the scope of application in practice. Especially DNNs act as black boxes due to their multilayer nonlinear structure. In this paper we introduce a novel methodology for interpreting generic multilayer neural networks by decomposing the network classification decision into contributions of its input elements. Although our focus is on image classification, the method is applicable to a broad set of input data, learning tasks and network architectures. Our method called deep Taylor decomposition efficiently utilizes the structure of the network by backpropagating the explanations from the output to the input layer. We evaluate the proposed method empirically on the MNIST and ILSVRC data sets.

KW - Deep neural networks

KW - Heatmapping

KW - Image recognition

KW - Relevance propagation

KW - Taylor decomposition

UR - http://www.scopus.com/inward/record.url?scp=85010676902&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85010676902&partnerID=8YFLogxK

U2 - 10.1016/j.patcog.2016.11.008

DO - 10.1016/j.patcog.2016.11.008

M3 - Article

VL - 65

SP - 211

EP - 222

JO - Pattern Recognition

JF - Pattern Recognition

SN - 0031-3203

ER -