Reproduce “The BatchNorm Illusion: Diagnosing Normalization Artifacts in Machine Unlearning Evaluation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Aaryaman Kalani, Murari Mandal, Dhruv Kumar, Mohan Kankanhalli, Yash Sinha
An analysis of batch-normalization state reveals that low forget accuracy in machine-unlearning evaluations can result from stale or corrupted running statistics rather than removal of target information. Recalibrating BN on retain data without changing model weights substantially restores forgotten-class performance in several methods, including a reported increase from 18.9% to 97.2% for BadTeacher, while improving the alignment of forget metrics with relearning and probing tests.
Key insights
- BN state can suppress apparent forgotten-class performance without removing the corresponding information from model weights or representations; recalibration therefore acts as a diagnostic for measurement bias.
- A small recalibration gap is ambiguous: it can indicate genuine forgetting, severe utility degradation, or accidental BN realignment during fine-tuning, so it must be interpreted alongside retain accuracy and leakage tests.
- The artifact is specific to normalization schemes with persistent running statistics: GroupNorm and LayerNorm eliminate the tested effect, whereas BN-based models can be vulnerable to recovery using only unlabeled images.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.