A preliminary study on class-incremental learning with repetition stream
Đang tải...
Năm xuất bản
2023
Tác giả
Nhan đề tạp chí
ISSN
Nhan đề tập
Nhà xuất bản
Giao thông vận tải
Tóm tắt
This paper presents a preliminary study on class - incremental learning with a repetition stream, which is a variant of continual learning that involves incrementally adding new classes to a classification model while revisiting and retraining on previously learned classes. The study evaluates the effectiveness of two different class-incremental learning algorithms using the repetition stream approach on the CIFAR-100 dataset. The study also investigates the impact of various factors such as the number of repetitions, the order of presentation in the class, and the size of the incremental task on the performance of the repetition stream approach. Our results indicated that the combination of EWC and LwF outperformed the individual algorithms.