arxivcs.LG2026-07-23
Gradient Concentration, Not Weight Saliency, Explains Representation-Level Class Unlearning
Billel Habbati, Alessio Merlo, Luca Verderame, Meriem Guerar
Machine unlearning aims to remove the influence of specific training data while preserving model utility. Many state-of-the-art approaches pursue this goal by restricting the forgetting update to a subset of parameters selected through gradient-based saliency. Although such metho…