Explaining nonlinear classification decisions with deep Taylor decomposition

PATTERN RECOGNITION

卷:65

Article

Montavon, Gregoire¹  Lapuschkin, Sebastian²  Binder, Alexander³  Samek, Wojciech²  Mueller, Klaus-Robert^1,4 

[1] Tech Univ Berlin, Dept Elect Engn & Comp Sci, Marchstr 23, D-10587 Berlin, Germany

[2] Fraunhofer Heinrich Hertz Inst, Dept Video Coding & Analyt, Einsteinufer 37, D-10587 Berlin, Germany

[3] Singapore Univ Technol & Design, Informat Syst Technol & Design, 8 Somapah Rd,Bldg 1, Singapore 487372, Singapore

[4] Korea Univ, Dept Brain & Cognit Engn, Anam Dong 5ga, Seoul 136713, South Korea

关键词: Deep neural networks;    Heaimapping;    Taylor decomposition;    Relevance propagation;    Image recognition;   

DOI  :  10.1016/j.patcog.2016.11.008

来源: Elsevier

【 摘 要 】

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.

【 授权许可】

Free