Interpreting Deep Visual Representations via Network Dissection
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 in...
| Main Authors: | , , , |
|---|---|
| Other Authors: | |
| Format: | Article |
| Language: | English |
| Published: |
Institute of Electrical and Electronics Engineers
2019
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/1721.1/122817 |
| _version_ | 1826210275788849152 |
|---|---|
| author | Zhou, Bolei Bau, David Oliva, Aude Torralba, Antonio |
| author2 | Massachusetts Institute of Technology. Computer Science and Artificial Intelligence Laboratory |
| author_facet | Massachusetts Institute of Technology. Computer Science and Artificial Intelligence Laboratory Zhou, Bolei Bau, David Oliva, Aude Torralba, Antonio |
| author_sort | Zhou, Bolei |
| collection | MIT |
| description | 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 |
| first_indexed | 2024-09-23T14:47:13Z |
| format | Article |
| id | mit-1721.1/122817 |
| institution | Massachusetts Institute of Technology |
| language | English |
| last_indexed | 2024-09-23T14:47:13Z |
| publishDate | 2019 |
| publisher | Institute of Electrical and Electronics Engineers |
| record_format | dspace |
| spelling | mit-1721.1/1228172022-10-01T22:27:43Z Interpreting Deep Visual Representations via Network Dissection Zhou, Bolei Bau, David Oliva, Aude Torralba, Antonio Massachusetts Institute of Technology. Computer Science and Artificial Intelligence Laboratory Computational Theory and Mathematics Software Applied Mathematics Artificial Intelligence Computer Vision and Pattern Recognition 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 United States. Defense Advanced Research Projects Agency (FA8750-18-C-0004) National Science Foundation (U.S.)(Grant 1524817) National Science Foundation (U.S.)(Grant 1532591) United States. Office of Naval Research (Grant N00014-16-1-3116) Massachusetts Institute of Technology. Computer Science and Artificial Intelligence Laboratory Google (Firm) Amazon.com (Firm) NVIDIA Corporation Facebook (Firm) 2019-11-11T18:43:22Z 2019-11-11T18:43:22Z 2019-09-01 2019-07-11T17:07:04Z Article http://purl.org/eprint/type/JournalArticle 0162-8828 2160-9292 1939-3539 https://hdl.handle.net/1721.1/122817 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 en http://dx.doi.org/10.1109/tpami.2018.2858759 IEEE Transactions on Pattern Analysis and Machine Intelligence Creative Commons Attribution-Noncommercial-Share Alike http://creativecommons.org/licenses/by-nc-sa/4.0/ application/pdf Institute of Electrical and Electronics Engineers arXiv |
| spellingShingle | Computational Theory and Mathematics Software Applied Mathematics Artificial Intelligence Computer Vision and Pattern Recognition Zhou, Bolei Bau, David Oliva, Aude Torralba, Antonio Interpreting Deep Visual Representations via Network Dissection |
| title | Interpreting Deep Visual Representations via Network Dissection |
| title_full | Interpreting Deep Visual Representations via Network Dissection |
| title_fullStr | Interpreting Deep Visual Representations via Network Dissection |
| title_full_unstemmed | Interpreting Deep Visual Representations via Network Dissection |
| title_short | Interpreting Deep Visual Representations via Network Dissection |
| title_sort | interpreting deep visual representations via network dissection |
| topic | Computational Theory and Mathematics Software Applied Mathematics Artificial Intelligence Computer Vision and Pattern Recognition |
| url | https://hdl.handle.net/1721.1/122817 |
| work_keys_str_mv | AT zhoubolei interpretingdeepvisualrepresentationsvianetworkdissection AT baudavid interpretingdeepvisualrepresentationsvianetworkdissection AT olivaaude interpretingdeepvisualrepresentationsvianetworkdissection AT torralbaantonio interpretingdeepvisualrepresentationsvianetworkdissection |