|
📍 LIEU / PLACE
Hybride - CRM, Salle / Room 6214, Pavillon André Aisenstadt
✒️ TITRE / TITLE
Statistical Frameworks for Trustworthy Machine Learning: Privacy, Uncertainty, and Online Inference
📄 RÉSUMÉ / ABSTRACT
Trustworthy machine learning requires rigorous privacy protection and valid uncertainty quantification, especially in modern streaming settings. We present two complementary advances.
First, we extend Gaussian Differential Privacy (GDP) to general Riemannian manifolds. Leveraging the Bishop-Gromov comparison theorem, we construct a Riemannian Gaussian mechanism based on geodesic distance and calibrate the privacy parameter $\mu$ to achieve GDP under bounded Ricci curvature. We provide practical calibration methods, an efficient procedure on one-dimensional manifolds and an MCMC-based algorithm on constant-curvature spaces. We demonstrate improved utility (e.g., on the sphere $S^d$) relative to Riemannian Laplace mechanisms.
Second, we develop online inference for smoothed quantile regression, introducing an incremental updating estimator for low-dimensional models and an online debiased lasso for high-dimensional sparse settings. The procedures use only current data and compact history summaries, correct online approximation error, and deliver asymptotically valid confidence intervals and tests. Simulations and real data analyses (e.g., bike-sharing demand and index-fund data) illustrate reliability and scalability.
Together, these results provide privacy-preserving data access and statistically sound, streaming-ready inference, core ingredients of trustworthy machine learning.
🍷 Une réception vins et fromages suivra / A wine and cheese reception will follow.
|