I think parent is getting at the idea that dnn learns abstract representations of the data in whatever way is useful for the mode objective function. But I agre

I think parent is getting at the idea that dnn learns abstract representations of the data in whatever way is useful for the mode objective function. But I agree, the article still seems to present a reasonable argument