Gradient Concentration, Not Weight Saliency, Explains Representation-Level Class Unlearning

2026-07-23

AI Summary

This study investigates whether the contribution of weight saliency selection to representation-level forgetting in machine unlearning has been overestimated. By controlling computational budgets, the authors conduct the first ablation study of saliency masking mechanisms, comparing saliency-based masks, random masks, and unconstrained updates on ResNet-18. Using representation evaluation methods—including linear probing, prototype recovery, and inter-layer centered kernel alignment (CKA)—they find no significant differences among the three strategies in terms of representational recoverability. Notably, 92% of gradient energy concentrates in the final few layers, and saliency masks exhibit extremely low class specificity (0.09–0.11), suggesting that gradient concentration—not weight selection—primarily drives representational forgetting.

Abstract

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 methods are widely adopted, the actual contribution of saliency-based weight selection to representation-level forgetting remains unclear. In this work, we perform the first controlled ablation of the saliency masking mechanism used by SalUn. Using a matched-compute experimental design on CIFAR-10 and CIFAR-100 with ResNet-18, we compare saliency-based masking against random masks of equal sparsity and unconstrained updates, while keeping the unlearning objective, optimization schedule, and computational budget fixed. Across multiple representation-level evaluations, including linear probing, prototype recovery, and layer-wise CKA, the three configurations exhibit statistically equivalent representation-level recoverability. We find that forget gradients are strongly concentrated in the final network layers (approximately 92% of the squared gradient energy on CIFAR-10) before any mask is applied, causing all masking strategies to operate within the same representational subspace. Furthermore, saliency masks show limited class specificity (specificity index 0.09-0.11), selecting highly overlapping parameter subsets across different forget classes. Our findings suggest that, in the studied setting, representation-level forgetting is primarily governed by gradient concentration and representation geometry rather than by the specific identity of saliency-selected weights. More broadly, the results support a growing body of evidence indicating that effective representation-level unlearning requires objectives that act directly on latent representations rather than on increasingly sophisticated weight-selection strategies.

Problem

• machine unlearning
• gradient concentration
• weight saliency
• representation-level forgetting
• saliency masking

Innovation

• gradient concentration
• representation-level unlearning
• saliency masking
• machine unlearning
• ablation study