The Utility and Complexity of in- and out-of-Distribution Machine Unlearning
The paper investigates how to selectively remove data from trained machine learning models (unlearning) while maintaining model performance. It explores two scenarios: removing data similar to the training set (in-distribution) and removing significantly different data (out-of-distribution). The authors develop new methods, including a robust gradient descent variant, proving that unlearning can be done efficiently without losing too much accuracy.