神经网络的可解释性综述!
共 4665字,需浏览 10分钟
· 2021-07-29
地址|https://zhuanlan.zhihu.com/p/368755357
本文仅作为学术分享,著作权归属原作者,侵删
本文以 A Survey on Neural Network Interpretability 读后感为主,加上自身的补充,浅谈神经网络的可解释性。
本文按照以下的章节进行组织:
人工智能可解释性的背景意义 神经网络可解释性的分类 总结
01
解释(Explanations),是指需要用某种语言来描述和注解
可解释的边界(Explainable Boundary),是指可解释性能够提供解释的程度
可理解的术语(Understandable Terms),是指构成解释的基本单元
高可靠性的要求
伦理/法规的要求
作为其他科学研究的工具
02
![](https://filescdn.proginn.com/b66cba5545dc1ae4ae72a42fe32216ae/7acb5de131054bdbbd6667e55e123b12.webp)
![](https://filescdn.proginn.com/9e6f10717545e5ac1419b4bad6c600b2/6fb25b671b64f62e0c17bdcbfb33a343.webp)
![](https://filescdn.proginn.com/011572d8cde8c1b970232330b95d6c8a/5ca3b9ca40b433fc8774ecce5ce0b4ed.webp)
![](https://filescdn.proginn.com/7c7a238e3f63a58f85ac89bedd349267/2643904f52098d7c913e78f08a2e9aee.webp)
![](https://filescdn.proginn.com/b0d824ac5f3d9297febd92cac27334b7/61adeea2e6ce61c2457a864a06fa608b.webp)
![](https://filescdn.proginn.com/1239b882286063d15e472e5b421af930/a24340601e554b07963b1eb057d379f1.webp)
![](https://filescdn.proginn.com/ae51e46d71d3b4c74e37666355816136/6d0ca848f133a2c5c855dc9968943c29.webp)
![](https://filescdn.proginn.com/ddac91cc7066e159eefb5b23e1768451/cb42e435ea3f760436a85fc5a8b7153b.webp)
![](https://filescdn.proginn.com/ee05ff4f1fbfb2307603802c163b5338/d9a62d99be61f4dfc4681d7a0d11058a.webp)
featuremap_layout
![](https://filescdn.proginn.com/2fac79ea1519ff8b58a7329d05479dbc/16f31f440b9f585c26ce38e72406490f.webp)
![](https://filescdn.proginn.com/328e83d3fd7e4f298a4b751027b0a3bf/db21b4ee5e08bcaaa47c653d51b8facf.webp)
![](https://filescdn.proginn.com/0fb07e2bb2865623eb15641aa53148e7/70eebf63e1d50f8a14c7da20350ce9ed.webp)
![](https://filescdn.proginn.com/52884f1f1b8ec1bd750cac91f5e0dee4/9f152b8b740bde57ea28936d22a36360.webp)
![](https://filescdn.proginn.com/ed4ca3838f27688c1449299c7c8193eb/a42b67aeff3e4e17c3975eb1db14ae8b.webp)
![](https://filescdn.proginn.com/084b4836e1352ecb546f98973694358a/09e75f603faad7a44874dbe0abb31b80.webp)
![](https://filescdn.proginn.com/be5208d88de5082baa5429dabb19f6c1/c03e9c83ba6f941a8bd4734361357630.webp)
![](https://filescdn.proginn.com/9cd81de229e4e57cba0bb8cf676345f0/649fbe7f9c1ff5d78da9486496f96bd4.webp)
![](https://filescdn.proginn.com/f1b4569052efaf4ebc57400f34a785df/3b66529e4e035828c43b95553a2d5b11.webp)
03
![](https://filescdn.proginn.com/5a70ab168558dfee902ae270b01c17c4/49cbccbd8ffdaa34a59e7d303e4c1e3b.webp)
![](https://filescdn.proginn.com/a40de7d7b10a4ce7626a3d81e839df77/1f693a0347d82a86382014b459f4b55f.webp)
参考
^“Extracting Decision Trees From Trained Neural Networks”. SIGKDD. July 23-26,2002 https://dl.acm.org/doi/10.1145/775047.775113
^ M. Wu, S. Parbhoo, M. C. Hughes, R. Kindle, L. A. Celi, M. Zazzi, V. Roth, and F. Doshi-Velez, “Regional tree regularization for interpretability in deep neural networks.” in AAAI, 2020, pp. 6413–6421. https://arxiv.org/abs/1908.04494
^K. Simonyan, A. Vedaldi, and A. Zisserman, “Deep inside convolutional networks: Visualising image classification models and saliency maps,” arXiv preprint arXiv:1312.6034, 2013.
^Q. Zhang, Y. Nian Wu, and S.-C. Zhu, “Interpretable convolutional neural networks,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018.
^P. W. Koh and P. Liang, “Understanding black-box predictions via influence functions,” in Proceedings of the 34th International Conference on Machine Learning-Volume 70, 2017.
^M. T. Ribeiro, S. Singh, and C. Guestrin, “Why should i trust you?: Explaining the predictions of any classifier,” in Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, 2016.
^M. Wojtas and K. Chen, “Feature importance ranking for deep learning,” Advances in Neural Information Processing Systems, vol. 33, 2020.
^Open Domain Dialogue Generation with Latent Images Z Yang, W Wu, H Hu, C Xu, Z Li - arXiv preprint arXiv:2004.01981, 2020 - arxiv.org https://arxiv.org/abs/2004.01981
——The End——
读者,你好!我们建立了微信学习交流群,欢迎扫码进群讨论!
微商、广告无关人员请绕道!谢谢合作!
![](https://filescdn.proginn.com/5d25ccd18376fa16443622a7577d36f2/107d6b977b73f3fac8e0a1a384e612d3.webp)