The paper introduces an information-theoretic metric for evaluating machine unlearning models, which helps assess how effectively models forget specific data while preserving overall performance.
The paper proposes a method to improve online continual learning by addressing underfitting from limited training, enhancing the model's adaptability to streaming data.
Award
2nd Place in Class-Incremental with Repetition (CIR) using Unlabeled Data - 5th CLVISION workshop @ CVPR 2024