Statistical Foundations of Machine Learning: A Review
An 18-page review connecting statistical theory with published case studies to argue that many AI fairness failures are statistical rather than software bugs.
Topic
Calibration, the bias–variance trade-off, sampling theory, and fairness impossibility results in machine learning.
Motivation
The paper argues that many AI fairness failures are statistical phenomena rather than software bugs, and explains why using existing theory and well-documented cases.
Approach
It synthesizes published work, including the Gender Shades audit, reporting on Amazon’s hiring tool, the COMPAS analysis, and the impossibility results of Chouldechova and of Kleinberg, Mullainathan and Raghavan.
Its one original element is a small bias–variance simulation in Python with scikit-learn, included as a figure for illustration.
Limitations
- A review: it explains and connects other people’s work and makes no new theorem, method, dataset or empirical result.
- The simulation is illustrative, not a novel empirical contribution.