Unlearning with partial label learning

Abstract

Machine Unlearning describes the challenge of forgetting data points that were used for an initial training of a machine learning model. Data privacy concerns as well as safety of sensitive learning data are the driving motivation for the emergence of this field. The special case of class unlearning is a challenge, as an entire class is to be unlearned without affecting the accuracy of potentially very similar other classes. We propose a novel method for class unlearning that is robust, efficient and can be applied without having access to the full initial training data. The approach is based on disambiguation-free partial label learning and can be understood as a stabilized version of gradient ascent. Furthermore, we show how this approach can be applied to training data with negative quasiprobabilities which is a problem encountered in high energy physics.