Pre or Post-Softmax Scores in Gradient-based Attribution Methods, What is Best?

Gradient based attribution methods for neural networks working as classifiers use gradients of network scores. Here we discuss the practical differences between using gradients of pre-softmax scores versus post-softmax scores, and their respective advantages and disadvantages.

cs.LG cs.AI Neural Networks and Applications Stochastic Gradient Optimization Techniques Machine Learning and ELM Computer Science

View graph Source

Paper

References (8)

01Visual Explanations from Deep Networks via Riemann-Stieltjes Integrated Gradient-based LocalizationMirtha Lucas, M. Lerma, Jacob D. Furst et al.2022 · International Symposium on Visual Computing · 22 citations In Library

02Grad-CAM++ is Equivalent to Grad-CAM With Positive GradientsMiguel A. Lerma, Mirtha Lucas2022 · 24th Irish Machine Vision and Image Processing Conference · 19 citations In Library

03Grad-CAM++: Generalized Gradient-Based Visual Explanations for Deep Convolutional NetworksAditya Chattopadhyay, Anirban Sarkar, Prantik Howlader et al.2017 · IEEE Workshop/Winter Conference on Applications of Computer Vision · 3.3k citations In Library

04Axiomatic Attribution for Deep NetworksMukund Sundararajan, Ankur Taly, Qiqi Yan2017 · International Conference on Machine Learning · 8.4k citations In Library

05Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based LocalizationRamprasaath R. Selvaraju, Abhishek Das, Ramakrishna Vedantam et al.2016 · International Journal of Computer Vision · 28k citations In Library

06Very Deep Convolutional Networks for Large-Scale Image RecognitionK. Simonyan, Andrew Zisserman2014 · International Conference on Learning Representations · 113k citations In Library

07Grad - CAM implemenation in tensorflow 2 . x . [ Online ]2021 · Proceedings of the 34 th International Conference on Machine Learning , ser . Proceedings of Machine Learning Research

08Gradient-Based Attribution MethodsMarco Ancona, Enea Ceolini, Cengiz Öztireli et al.2019 · Explainable AI · 191 citations