Accuracy of NS, IF, RIF, and DRIF on Synthetic Gaussian Data

Read full text Summarise with VIRAT
Authors: Ittai Rubinstein
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

We generate a synthetic logistic-regression problem with n = 2^16 Gaussian covariates in dimension d = 512 and labels drawn from a well-specified model with ground-truth parameter θ* = e_1 = (1, 0, ..., 0). For each subset size k on the x-axis, up to half the dataset, we remove a uniformly random set of k training points and compare each method’s predicted leave-k-out parameter vector with the parameter obtained by full retraining. The y-axis reports error in the Mahalanobis norm induced by Σ = X^T X, i.e. the unnormalized second-moment matrix of the covariates. Across the full range of subset sizes, and especially for larger removals, NS and DRIF are substantially more accurate than IF and RIF, consistent with our theoretical analysis.

Fields of Study & Subjects

Machine Learning and Data Classification Advanced Statistical Methods and Models Stochastic Gradient Optimization Techniques

Publication details

Journal / source
Figshare
Publication date
31 March 2026
Language
EN