Add more methods to the feature importance page

[aperloff]

Based on a CMSTalk conversation there are several methods for feature importance we may wish to add.

Methods:

• Taylor expansion (usually up to second order) of the NN output function w.r.t. the input features. Can be done during (learn when a feature becomes important) or after training (learn the most important features). Already discussed during a journal club. This presentation contains a Jupyter example. The method is described in this paper.

Tools section:

• Several methods are encapsulated in the innvestigate package. Andrzej Novak recommends the "integrated gradients" method.
• The shap module can be used to obtain a ranking of the input features. The method is based on the Shapley values, which originally come from game theory and are computed by integrating-out sets of input features. This is somewhat similar to the "recursive feature elimination" which is already mentioned in the documentation.