Algorithmic Transparency via Quantitative Input Influence - Theory and Experiments with Learning Systems

Explore the concept of algorithmic transparency in machine learning systems through this IEEE Symposium on Security & Privacy presentation. Delve into Quantitative Input Influence (QII) measures, designed to capture the degree of influence inputs have on system outputs. Learn how these measures can be applied to create transparency reports, detect algorithmic discrimination, and explain individual and group decisions. Examine the challenges of correlated inputs, the importance of joint and marginal influence, and the use of game theory concepts like the Shapley value. Discover how QII measures can be made differentially private while maintaining accuracy. Gain insights into the practical applications and empirical validation of QII measures with standard machine learning algorithms, and understand their advantages over traditional associative measures in various scenarios.