Notes on “fairness of learning”

1.	We discussed how to define the fairness and what is the right notion for it
	*	Kamalika mentioned a law article to define the fairness. For instance, there are “disparate impact” and “disparate treatment”. The point is we sometimes ignore the disparate impact by focusing on practical treatment and do not think further for the impact.
	*	Is fairness another loss of learning? Percy thinks somehow such fairness exists due to certain reasons and maybe that’s another way to interpret the prediction


2.	Domain where fairness of learning should be considered
	*	for instance, it can be applied to protect attributes
	*	it involves privacy problems, since learning without fairness can emphasize certain features or directions that generate private information leakage
	*	predictions should be orthogonal with attributes that are sensitive 


3.	what’s the enumerated set of attributes that will lead to break the fairness
	*	it is important to define the “group” that will optimize the “impact” and try to control them and guarantee fairness


4.	How to balance the accuracy and fairness during learning, or should we choose one of them for certain applications?