Upweighting easy samples (low pre-trained loss) during fine-tuning reduces catastrophic forgetting without access to pre-training data, at a small cost in target performance.
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
-
Upweighting Easy Samples in Fine-Tuning Mitigates Forgetting
Upweighting easy samples (low pre-trained loss) during fine-tuning reduces catastrophic forgetting without access to pre-training data, at a small cost in target performance.