Under Adam, feeding a modified gradient into both moment accumulators cancels the intended continual-learning protection; feeding only the first moment preserves it.
Gradient projection for parameter-efficient continual learning.arXiv preprint arXiv:2405.13383
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Hidden Failure Modes of Gradient Modification under Adam in Continual Learning, and Adaptive Decoupled Moment Routing as a Repair
Under Adam, feeding a modified gradient into both moment accumulators cancels the intended continual-learning protection; feeding only the first moment preserves it.