When Unlearning Is Free: Leveraging Low Influence Points to Reduce Computational Costs
| Source: Apple ML Research
Tags: machine unlearning, Apple, data privacy, GDPR, influence functions
Apple and Harvard researchers propose a machine unlearning framework that skips training data points with negligible model influence, cutting unlearning computation by up to 50%—a practical step toward scalable compliance with GDPR right-to-be-forgotten requirements.
Details
Apple ML Research and Harvard researchers published a paper on efficient machine unlearning—the process of removing specific training data from already-deployed models. The core insight: not all data points in a forget set actually influence model behavior meaningfully. By using influence functions to identify and skip low-impact points before running unlearning, the framework achieves up to ~50% computational savings on real-world language and vision tasks. Current state-of-the-art unlearning methods treat all forget-set points equally, which is wasteful when many had negligible effect on model parameters. This work challenges that assumption with comparative analysis across both vision and language tasks, identifying subsets that can safely be bypassed without degrading unlearning quality. The practical stakes are significant: GDPR's right to be forgotten and similar regulations increasingly require AI systems to demonstrate the ability to remove specific individuals' data. More efficient unlearning makes compliance feasible at scale—particularly for companies operating large models where full retraining is prohibitively expensive. The paper is authored by Udi Wieder, Vitaly Feldman, Robert Fisher, and Anat Kleiman (Harvard); available at machinelearning.apple.com.