search

albermax / innvestigate

A toolbox to iNNvestigate neural networks' predictions!
Other
1.24k stars 235 forks source link

[Question] Input_t_gradient vs LRP-Z #303

Closed tszoldra closed 1 year ago

tszoldra commented 1 year ago

Is it expected that the output heatmap of input_t_gradient analyzer is always the same as that of LRP-Z? I get it in all my analyses of VGG16, and see the same at least visually in the example notebooks, eg. in examples/mnist_compare_methods.ipynb Zrzut ekranu z 2023-01-07 15-35-32

I think that either the two methods are indeed equivalent in these special cases or it is a bug. In any case, it would be interesting to know. Thanks for your help in advance!

tszoldra commented 1 year ago

Indeed, after looking at the literature [G. Montavon, A. Binder, S. Lapuschkin, W. Samek, and K.-R. M ̈uller. Layer-wise relevance propagation: An overview. In Explainable AI, volume 11700 of Lecture Notes in Computer Science, pages 193–209. Springer, 2019] and [Shrikumar, A., Greenside, P., Shcherbina, A., Kundaje, A.: Not just a black box: Learning important features through propagating activation differences. CoRR abs/1605.01713 (2016)] I understand that the two methods are equivalent when applied to the whole network.