Transformers can be trained with enforced spectral-norm constraints throughout training, but competitive accuracy requires an astronomical Lipschitz upper bound.
Notably, the official implementation of LipsFormer uses ϵ = 0
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
-
Training Transformers with Enforced Lipschitz Constants
Transformers can be trained with enforced spectral-norm constraints throughout training, but competitive accuracy requires an astronomical Lipschitz upper bound.