Explore fairness in representation learning through this insightful conference talk by Natalie Dullerud, an incoming PhD student at Stanford University. Delve into the evaluation and mitigation of bias in deep metric learning (DML), focusing on subgroup disparities. Examine the negative impact of imbalanced training data on minority subgroup performance in downstream tasks. Learn about the proposed fairness in non-balanced DML benchmark (finDML) and its three key properties: inter-class alignment, intra-class alignment, and uniformity. Discover how bias in DML representations propagates to common downstream classification tasks, even when training data is re-balanced. Understand the Partial Attribute De-correlation (PARADE) method, designed to reduce performance gaps between subgroups in both embedding space and downstream metrics. Gain insights into broader fairness metrics in representation learning and their potential applications across different domains.
Overview
Syllabus
Fairness in Representation Learning A study in evaluation and mitigation of bias via subgroup
Fairness in Machine Learning
Fairness in Representations: DML
Overview: Fairness in Deep Metric Learning
Intuition: Fairness in DML
Defining Fairness in DML
Experimental Design
Empirical Results: Bias Propagates
Bias Mitigation: Considerations
Bias Mitigation: An Initial Solution (PARADE)
Empirical Results in PARADE
Comparison with Oversampling
Limitations to PARADE
Fairness Improvements in Representations
Thank you for listening!
PARtial Attribute DE-correlation (PARADE)
Taught by
Stanford MedAI
