基于 Z 分数的 CNN 模型可解释性
计算机视觉与模式识别
2021-02-12 v1
摘要
本文探讨了神经网络(NNs)输出层与逻辑回归的相似性,以通过 Z 分数解释输入的重要性。所分析的网络是一个用于融合合成孔径雷达(SAR)与微波辐射计(MWR)数据以预测北极海冰的网络。通过该分析,发现 MWR 相对于 SAR 的重要性偏向于 MWR 分量。此外,由于模型在不同尺度上表示图像特征,这些特征的相对重要性也得到了分析。所提出的方法提供了一个简单且易用的框架来分析输出层分量,并可使用例如常见的 NN 可视化方法减少进一步分析的分量数量。
引用
@article{arxiv.2102.05874,
title = {Explainability in CNN Models By Means of Z-Scores},
author = {David Malmgren-Hansen and Allan Aasbjerg Nielsen and Leif Toudal Pedersen},
journal= {arXiv preprint arXiv:2102.05874},
year = {2021}
}
备注
Intended and accepted for the "Deep Learning Meets Earth Sciences: From Hybrid Modeling to Explainability" workshop at IGARSS 2020, but was redrawn due to authors being unable to participate when lockdown restrictions moved the conference days. The work was conducted 2019 under the Automated Sea Ice Products (ASIP) project funded by the Innovation Fund Denmark