TA-KAN
A replay-free continual learning framework based on Kolmogorov-Arnold Networks (KAN) that employs a task-adaptive classifier and geometry-guided out-of-distribution scoring to improve out-of-distribution detection performance while avoiding catastrophic forgetting.
Timeline (last 90 days)
TA-KAN is a framework for rehearsal-free continual learning scenarios, aiming to simultaneously reduce forgetting of old tasks and accurately identify out-of-distribution (OOD) samples.
TA-KAN designs a task-adaptive classifier by controlling activation locality on a per-task basis, in order to enhance the separability of ID and OOD samples and suppress recency bias.
The paper proposes a geometry-guided OOD score as a complement to TA-KAN's confidence; this score does not depend on any past ID data and is therefore compatible with the replay-free setting.
The paper reports that TA-KAN substantially improves OOD detection performance while maintaining strong ID classification performance, and achieves state-of-the-art results on the OOD-CL benchmark.
At the abstract level, the paper does not disclose the specific benchmark datasets, task sequence lengths, or list of compared methods, and the stability of KAN-type backbones on larger-scale task sequences still requires more experimental validation.
All Facts (5)
TA-KAN is a framework for rehearsal-free continual learning scenarios, aiming to simultaneously reduce forgetting of old tasks and accurately identify out-of-distribution (OOD) samples.
50%UnverifiedTA-KAN designs a task-adaptive classifier by controlling activation locality on a per-task basis, in order to enhance the separability of ID and OOD samples and suppress recency bias.
50%UnverifiedThe paper proposes a geometry-guided OOD score as a complement to TA-KAN's confidence; this score does not depend on any past ID data and is therefore compatible with the replay-free setting.
50%UnverifiedThe paper reports that TA-KAN substantially improves OOD detection performance while maintaining strong ID classification performance, and achieves state-of-the-art results on the OOD-CL benchmark.
50%UnverifiedAt the abstract level, the paper does not disclose the specific benchmark datasets, task sequence lengths, or list of compared methods, and the stability of KAN-type backbones on larger-scale task sequences still requires more experimental validation.
50%