adjustment for softmax about deeplift HOT 2 CLOSED

It's more like this:

Original feature contributions to the class logits:
class 0: {1, 1, 0, -1}
class 1: {1, 0, 1, -3}

Mean feature contributions across logits of all classes: {1, 0.5, 0.5, -2}

Mean-normalized feature contributions:
class 0: {0, 0.5, -0.5, 1}
class 1: {0, -0.5, 0.5, -1}

Does that make sense? The goal is to account for the fact that if a feature contributes equally to the logits of all softmax classes, that means it effectively contributes to none of the softmax classes. For example, the first feature is contributing "1" to both class 0 and class 1; that means it is effectively contributing to none of the classes (because the contribution would cancel out the numerator and denominator of the softmax).

from deeplift.

cystral clear now ... thank you so much ...

from deeplift.