Core Concepts
Task-Adaptive Saliency Supervision (TASS) mitigates saliency drift in Exemplar-free Class Incremental Learning, improving performance and reducing forgetting.
Stats
EFCIL aims to learn tasks sequentially without access to previous data.
TASS integrates boundary-guided saliency, low-level supervision, and saliency noise injection.
TASS outperforms state-of-the-art methods on CIFAR-100, Tiny-ImageNet, and ImageNet-Subset benchmarks.
Quotes
"Our experiments demonstrate that our method can better preserve saliency maps across tasks and achieve state-of-the-art results."
"TASS aims at keeping saliency focused on incrementally learned tasks while maintaining its plasticity and stability."