Random augmentation can trigger gradient collisions ("evil twins") that cause forgetting; selectively averaging drifted weights with a snapshot improves single-source domain generalization.
A data- augmentation is worth a thousand samples: Analytical mo- ments and sampling-free training
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
support 1representative citing papers
citing papers explorer
-
Dealing with the Evil Twins: Improving Random Augmentation by Addressing Catastrophic Forgetting of Diverse Augmentations
Random augmentation can trigger gradient collisions ("evil twins") that cause forgetting; selectively averaging drifted weights with a snapshot improves single-source domain generalization.