SEMU forgets target data by adding to each network layer a low-rank correction built from the top singular vectors of the forget-set gradient, training only the small correction matrix.
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
-
SEMU: Singular Value Decomposition for Efficient Machine Unlearning
SEMU forgets target data by adding to each network layer a low-rank correction built from the top singular vectors of the forget-set gradient, training only the small correction matrix.