The repository was written during the seminar Mathematical Foundations of Deep Learning topic 12: Deep Taylor Decomposition at LMU munich.
The authors or the original paper, which can be found online, propose a method to extend the explainability of deep neural networks with taylor decompositions. The here presented repositories delivers a summary of the given work in form of a short paper and presentation. Since Montavon et al. only present examples from the field of image recognition I will extend their experiment to a common data science classification problem, the titanic data set. I chose this particular data set in favor of other famous data sets, like the iris data set, henceI like the interpretability of the results, like the influence of the gender on the survival probability.
the folder Code contains all the code, sme experimental replica of the original algorithms as well as the titanic example. The folder paper contains the small 5 pages summary of the original paper.