Kurzkommentar |
Abstract:
The concepts of bias, fairness and transparency have been studied thoroughly in philosophical, social and cultural frameworks. However, as automated data analysis replaces human supervision and intuition in decision making, and the scale of the data analyzed becomes “big”, there is a growing need for incorporating these concepts into algorithmic decision-making frameworks [1, 2,3]. While these concepts are somewhat intuitive, quantifying them in algorithmic frameworks is a non-trivial task. For example, what does it mean for an algorithm to be fair or unbiased? Or, how do we quantify the extent to which an algorithm is discriminating against subjects of a protected class, and so on. In this seminar, we will try to answer some of these questions. We will start off by studying the definitions of these concepts and try to map them to algorithmic frameworks. Algorithmic way of quantifying these concepts exposes their underlying complexity. We will see that the philosophical definitions of these concepts allow for a large number of subjective solutions. For example, different humans can interpret the definition of impartiality (or lack of bias) in different ways. On the other hand, the algorithmic quantification of bias has to be very precise. In order to solve these problems, we will take a computational approach towards understanding these concepts. |