keyword
non-exemplar class-incremental learning
Non-exemplar class-incremental learning is a machine learning paradigm in which a model continuously learns to recognize new categories of data over time while retaining the ability to classify previously learned categories, without storing or replaying any past training samples. In this setting, a system must adapt to sequential tasks and expand its classification capabilities under strict privacy, storage, or computational constraints that prohibit saving historical exemplars. Because old data cannot be revisited during subsequent training stages, this approach focuses on preventing catastrophic forgetting and maintaining accurate decision boundaries between old and new classes by relying solely on current data, model parameter preservation, feature space regularization, or knowledge distillation.
1 item
