Proposal Description

S. Ejaz Ahmed will present new developments in post-shrinkage estimation for high-dimensional regression, addressing the bias that can occur when penalized methods fail to distinguish weak signals from null ones. His High-Dimensional Post-Shrinkage Estimation (HDPSE) approach improves predictive accuracy and interpretability when both strong and weak effects are present.
Yi Li will present a statistical inference framework for deep neural network models under generalized nonparametric regression settings with categorical and exponential family outcomes. The approach enables valid inference on DNN-estimated subject-specific means by allowing dependence between estimation errors and inputs, and uses an Ensemble Subsampling Method based on U-statistics and the Hoeffding decomposition to construct reliable confidence intervals. Simulation studies and an application to the eICU dataset illustrate accurate uncertainty quantification and patient-centric inference for clinical decision making.
Malka Gorfine will present FLEXI-Haz, a flexible deep neural network framework for survival analysis with a partially linear structure. The method combines a parametric linear component for covariates of primary interest with a nonparametric neural network component to model complex time-covariate interactions, without relying on the proportional hazard assumption. Strong theoretical guarantees support interpretable and efficient inference, and the approach is illustrated through simulations and real-data analyses.
Together, these talks highlight emerging directions in statistical learning and inference that connect methodological innovation with real-world data challenges. The presentations demonstrate how modern approaches can integrate prediction, uncertainty quantification, and interpretability within unified frameworks, including settings involving high-dimensional regression, survival analysis, and multi-cohort data integration. By combining tools from high-dimensional modeling, deep learning, and data integration, the speakers show how statistical learning can advance beyond pure prediction to yield deeper understanding of underlying mechanisms. This synthesis strengthens both the theoretical foundations and the practical impact of statistical methodology. The session will appeal to researchers interested in learning theory, high-dimensional inference, deep learning, and applications in biomedical and complex data settings.