A0963
Title: Double descent and benign overfitting for error in variables regression
Authors: Rishi Sonthalia - Boston College (United States) [presenting]
Abstract: Many prior works have used random matrix theory to understand the generalization risk for linear regression. The focus is on adding noise to both the output and the input. This change significantly affects the learning process. Using tools from random matrix theory, the generalization error is derived for low-dimensional data. Results are provided on covariate shifts and an astonishing phenomenon is noticed: double descent is obtained with under-parameterized models.
