Please note: This master’s thesis presentation will be given online.
Sushant
Agarwal, Master’s
candidate
David
R.
Cheriton
School
of
Computer
Science
The underlying message of this work is that the desiderata of fairness, interpretability, and privacy come at a cost, and it is necessary to make trade-offs between them. The first main result shows that there are inconsistencies between the requirements of fairness and interpretability of classifiers. We consider a formal framework to build simple classifiers as a means to attain interpretability. We show that each simple classifier is strictly improvable, in the sense that every simple classifier can be replaced by a more complex classifier that strictly improves both fairness and accuracy. The second main result considers the issue of compatibility between fairness and differential privacy of learning algorithms in settings where both might be a concern. We show an impossibility theorem which shows how even in simple classification settings, differential privacy and fairness are at odds with each other when we require a classifier with non-trivial accuracy.
To joint this master’s thesis presentation on Zoom, please go to https://zoom.us/j/288741325.
Meeting ID: 288 741 325