Promises and Pitfalls of Machine Learning for Education - Livestream

The widespread use of machine learning techniques in social settings remains controversial, with a variety of recent examples in education, healthcare, and criminal justice. In response, the machine learning community has produced a wide range of fairness measures that theoretically address different forms of algorithmic bias, but applying these measures in practice under noisy data or modern privacy requirements is no longer so theoretically clean. The first part of this talk will cover new methods applying robust optimization to handle fairness constraints under noisy protected group information. But fairness constraints are only part of the story - the second part of this talk will expand the picture of positive societal impact to a broader question of how ML can better support real world societal principles and goals. Using the education domain as a case study, we examine whether the stated or implied societal objectives of papers from highly-regarded ML conferences are aligned with the ML problem formulation, objectives, and interpretation of results. Through the lens of interviews with education domain experts, we expand the view of the ML life cycle to include a deeper dive into problem formulation and the translation from predictions to interventions.

Speaker: Serena Wang, UC Berkeley