Interpreting Deep Visual Representations via Network Dissection
Name
1711.05611.pdf
Description
Accepted version
Size
8.22 MB
Format
Adobe PDF
Checksum (MD5)
40f16e053b547d0ee4318f191262bf99
Author(s)
Zhou, Bolei
Bau, David
Oliva, Aude
Torralba, Antonio
Date Issued
September 1, 2019
Journal
IEEE Transactions on Pattern Analysis and Machine Intelligence
Publisher
Institute of Electrical and Electronics Engineers
Citation
Zhou, Bolei et. al. "Interpreting Deep Visual Representations via Network Dissection." Transactions on Pattern Analysis and Machine Intelligence, vol. 41, no. 9 (September 2019): pp. 2131-2145 © Institute of Electrical and Electronics Engineers 2019
Version
Author's final manuscript
Abstract
The success of recent deep convolutional neural networks (CNNs) depends on learning hidden representations that can summarize the important factors of variation behind the data. In this work, we describe Network Dissection, a method that interprets networks by providing meaningful labels to their individual units. The proposed method quantifies the interpretability of CNN representations by evaluating the alignment between individual hidden units and visual semantic concepts. By identifying the best alignments, units are given interpretable labels ranging from colors, materials, textures, parts, objects and scenes. The method reveals that deep representations are more transparent and interpretable than they would be under a random equivalently powerful basis. We apply our approach to interpret and compare the latent representations of several network architectures trained to solve a wide range of supervised and self-supervised tasks. We then examine factors affecting the network interpretability such as the number of the training iterations, regularizations, different initialization parameters, as well as networks depth and width. Finally we show that the interpreted units can be used to provide explicit explanations of a given CNN prediction for an image. Our results highlight that interpretability is an important property of deep neural networks that provides new insights into what hierarchical structures can learn. Keywords: Convolutional neural networks; Network interpretability; Visual recognition; Interpretable machine learning; Visualization; Detectors; Training; Image color analysis; Task analysis; Image segmentation; Semantics
Subjects
Computational Theory and Mathematics
Software
Applied Mathematics
Artificial Intelligence
Computer Vision and Pattern Recognition
MIT Department
Massachusetts Institute of Technology. Computer Science and Artificial Intelligence Laboratory
Terms of Use
Creative Commons Attribution-Noncommercial-Share Alike
Persistent DSpace Link
DOI of Published Version
https://doi.org/10.1109/tpami.2018.2858759
