How Hard Can It Be? Hardness-Aware Multi-Objective Unlearning
Two new research papers propose novel methods for machine unlearning, a process that removes specific data's influence from trained models. The first paper, "How Hard Can It Be? Hardness-Aware Multi-Objective Unlearning," introduces HAMU, which quantifies the difficulty of unlearning based on data similarity and guarantees specified improvements in forget quality while minimizing utility loss. The second paper, "Multi-Objective Reference-Aligned Machine Unlearning," presents RAUL, a framework that aligns forgotten sample predictions with a reference distribution to constrain forgetting and reduce conflicts with retention, aiming to minimize the gap compared to full retraining. AI
IMPACT These new unlearning techniques could improve data privacy and model management by offering more controlled ways to remove specific data influences.