Question about corrective term

[yinwangsong]

Hello! Thanks for your wonderful work!

I am not an expert of ML theory. I have a little question that, in detail, is the corrective term applied to gradient x input by simply adding or subtracting it?

Although it may be trivial but I wondered why the error is decreased given that gradient x input might be either larger or smaller than groundtruth attribution scores..?