Disentangling Influence: Using Disentangled Representations to Audit Model Predictions

Explore the concept of disentangled representations in auditing model predictions through this 27-minute lecture by Charlie Marx from Haverford College. Delve into recent developments in fairness research, examining direct and indirect influence, independent settings, and supervised learning processes. Understand the driving methodology behind disentangled representations, including independent factors, group actions, and world states. Investigate experimental results, focusing on reconstruction and prediction errors, feature audits, and shape-scale-orientation relationships. Analyze the limitations of this approach and engage with thought-provoking questions in the field of machine learning fairness.