PREREQUISITES:
Knowledge of statistical methods such as linear regression and logistic regression.
DESCRIPTION:
Testing independence and conditional independence is fundamental for many statistical and machine learning methods. In this talk, we present a binary-expansion- based framework for scalable and interpretable dependence testing. We first introduce a family of adaptive, distribution-free independence tests for multivariate random vectors using binary expansion coefficients. By reformulating the tests as U-statistics with explicit kernel representations, the proposed methods enable efficient computation while adaptively capturing complex dependence structures. Building on this framework, we propose DeepBET for testing conditional independence in the presence of highdimensional confounding variables. Deep neural networks are used to estimate conditional mean structures, and binary expansion tests are applied to residuals to detect remaining dependence. Despite the high-dimensional and nonparametric setting, the resulting test achieves root-n convergence. Simulations and real-data applications demonstrate strong power, accurate type I error control, and improved interpretability in complex nonlinear settings.
SPONSOR ACKNOWLEDGEMENT:
Sponsored by OSCTR BERD Core—see the attachment for the specifics
FORMAT:
Zoom