ART shows that many shortcut-mitigation and association-unlearning methods leave label-attribute associations that can be reactivated by amplifying residual feature directions under the original classifier head.
In: International Conference on Learning Representations (2024)
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Association Restoration Test: Revealing Restorable Shortcuts after Unlearning
ART shows that many shortcut-mitigation and association-unlearning methods leave label-attribute associations that can be reactivated by amplifying residual feature directions under the original classifier head.