A1458
Title: Algorithmicinference in proportional high dimensions
Authors: Xiaocong Xu - University of Southern California (United States) [presenting]
Abstract: Many modern learning problems are studied in a proportional high-dimensional regime, where the feature dimension is not negligible compared to the sample size. An algorithmic perspective on this regime views gradient descent and proximal gradient iterates as statistical estimators along the optimization path, rather than focusing only on the final empirical risk minimizer. Results characterizing the distribution of these iterates show how this characterization can be used to construct data-driven estimates of generalization error and debiased iterates for statistical inference, including in settings beyond linear regression. Simulations illustrate how tuning for prediction and tuning for inference may differ along the optimization trajectory.
