Uniform Herding Improves Class-Incremental Learning Accuracy
A recent preprint on arXiv (2608.13061) presents Uniform Herding, a novel technique for class-incremental learning that tackles the issue of maintaining older classes amidst evolving feature representations. This method employs a limited active exemplar set for the observed classes and utilizes a constrained candidate pool to update exemplars in the current representation. In tests conducted on CIFAR-100 involving ten class-incremental tasks, Uniform Herding, with a ResNet-18 backbone, an active budget of 2,000 exemplars, a retrieval budget of 64, and three seeds, achieved an average accuracy of 44.00% ± 0.51% and a forgetting rate of 17.22% ± 0.43%. In contrast, iCaRL recorded an accuracy of 42.33% ± 1.20% and a forgetting measure of 24.87% ± 1.11%. The research indicated that replacing NME or herding within the Uniform Herding framework led to decreased final accuracy, while the removal of distillation resulted in increased forgetting. Adjusting the retrieval budget had less impact than modifying the active budget. This paper has been announced as a new submission on arXiv.
Key facts
- Uniform Herding is a new method for class-incremental learning.
- It allocates a bounded active exemplar set across observed classes and uses a bounded candidate pool to refresh exemplars.
- Experiments on CIFAR-100 with ten class-incremental tasks, ResNet-18 backbone, active budget M=2,000, retrieval budget b=64, and three seeds.
- Uniform Herding achieved 44.00 ± 0.51% final average accuracy and 17.22 ± 0.43% forgetting.
- iCaRL achieved 42.33 ± 1.20% accuracy and 24.87 ± 1.11% forgetting.
- Replacing NME or herding with alternatives decreased final accuracy.
- Removing distillation increased forgetting.
- Changing retrieval budget had smaller effect than changing active budget.
- Paper is on arXiv with ID 2608.13061.
Entities
Institutions
- arXiv