Algorithmic fairness is an important and widely discussed topic in machine learning. However, it is often unclear what “fair” means or whether different stakeholders share the same understanding of what a fair algorithm should do. This talk will discuss various definitions of fairness and the philosophical stances that lead to different definitions, including concerns about discrimination against groups versus individuals and concerns about distributive versus procedural fairness. Similarly, it is also ambiguous at times what a transparent, explainable, or interpretable machine learning algorithm is. This talk will address the differences between interpreting a model and explaining an uninterpretable model, and highlight the value (and risk) of the etiological nature of transparent AI. Finally, the talk will cover some methods for improving fairness and transparency with respect to specific definitions.
And here are a few relevant readings of possible interest: