Abstract: Machine learning algorithms are getting used increasingly in making decisions in domains with social consequences. This leads to a natural concern about the biases the algorithms might learn from possibly biased historical data. In this talk, we would first discuss some examples that motivate the kind of properties a fair machine learning algorithms must possess. We would then discuss some of the popular definitions of fairness in machine learning and review some of the properties such definitions should have. The talk would be minimally technical and there are no prerequisites.