In-Depth Hyperparameter Selection For Layer-Wise Relevance Propagation
Contributors
- 1. TU Berlin
Description
Master's Thesis at TU Berlin's ML/IDA Group headed by Prof. Dr. Klaus-Robert Müller.
Abstract.
Our expectations of Explainable AI have grown together with its popularity. So far, the interpretability technique of Layer-Wise Relevance Propagation (LRP) has been adopted with mostly qualitative evaluation of its rules. Therefore, a quantitative and qualitative evaluation of LRP rules is conducted to determine which hyperparameters provide the best scoring heatmaps according to the Pixel-Flipping and Area Under the Curve evaluation framework. It can be concluded from the experiment results that the choice of evaluation metrics and visualization of heatmaps has a significant impact on explanations. Additionally, due to the inherent subjectivity of visual explanations the requirements should be defined on a case-by-case basis.
Files
master-thesis.pdf
Files
(16.3 MB)
Name | Size | Download all |
---|---|---|
md5:da67a571bb864e841463e807ea6484ba
|
16.3 MB | Preview Download |
Additional details
Related works
- Cites
- Software: 10.5281/zenodo.6821295 (DOI)