Back to papers
arxiv8.0 / 10

Bayesian Influence Functions for Hessian-Free Data Attribution

Abstract

Addresses challenges of classical influence functions in deep neural networks with non-invertible Hessians and high-dimensional parameter spaces

Research area

model robustnessrobustness to domain shifts
Published
Source
arxiv
Org
Timaeus
View paper
Sign in to read and join the discussion.