Apple proposes unlearning method that cuts computational cost by up to 50% by focusing on low-influence training points

Apple’s Machine Learning Research team proposes an unlearning framework that reduces computational costs by up to approximately 50% on real-world empirical examples. The method leverages influence functions to identify subsets of training data with negligible impact on model outputs, challenging the conventional approach of treating all points in the forget set equally.

The authors—Udi Wieder, Vitaly Feldman, Robert Fisher, and Anat Kleiman—argue that not all training points require removal during unlearning. By focusing on low-influence points, the framework reduces the size of datasets before unlearning, leading to significant computational savings.

The work includes a comparative analysis of influence functions across language and vision tasks, demonstrating the method’s applicability beyond a single domain. The authors note that this approach could make machine unlearning more efficient and scalable, particularly for privacy-sensitive applications where unlearning specific data points is required.

The research was published in August 2026 and builds on prior work in data privacy and machine learning, including a related reading on subspace recovery from heterogeneous data with non-isotropic noise.