I'm still trying to understand "algorithmic bias is an algorithm problem". Does that mean, if I select L1, I'm training a network which is less accurate against examples that have rare features, because the weights on those features will be forced to 0?
I just want to understand if that's the rough idea or I'm far from the point. If my thinking is approximately correct, then I have a series of further questions and comments, but I think I must be understanding what shooker is saying.
I just want to understand if that's the rough idea or I'm far from the point. If my thinking is approximately correct, then I have a series of further questions and comments, but I think I must be understanding what shooker is saying.