MoTE, which combines per-task adapters with task-scope expert filtering and confidence-weighted feature fusion, reports state-of-the-art average accuracy for exemplar-free class-incremental learning on CIFAR100, CUB, ImageNet-A, ImageNet-R, and VTAB.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
MoTE: Mixture of Task-specific Experts for Pre-Trained ModelBased Class-incremental Learning
MoTE, which combines per-task adapters with task-scope expert filtering and confidence-weighted feature fusion, reports state-of-the-art average accuracy for exemplar-free class-incremental learning on CIFAR100, CUB, ImageNet-A, ImageNet-R, and VTAB.