Fairness and Robustness in Machine Learning – A Formal Methods Perspective - Aditya Nori, Microsoft

Explore a 34-minute conference talk on fairness and robustness in machine learning from a formal methods perspective. Delve into the imperative need to investigate fairness and bias in decision-making programs as algorithmic decisions become more prevalent and sensitive. Learn about encoding formal definitions of fairness as probabilistic program properties and discover a novel technique for verifying these properties across a wide range of decision-making programs. Examine FairSquare, the first verification tool for automatically certifying a program's fairness, and understand its evaluation on various decision-making programs. Gain insights into the intersection of logic and learning, exploring how formal reasoning and statistical approaches can be combined to address complex problems in machine learning fairness and robustness.

Introduction
New programming language challenges
How does one formalize the notion of fairness
Algorithmic decision making
Questions
Question
Population model
Symbolic execution
Invariants
Triangle example
Hyper rectangular decomposition
Subsampling hyper rectangles
Challenges with sampling
Ideal solution
Approximate density
Properties
Proofs
Summary