Conservation laws for gradient-flow training of conv ResNets and Transformers are characterized for several building blocks, and deep network block laws reduce to laws of isolated blocks.
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
-
Transformative or Conservative? Conservation laws for ResNets and Transformers
Conservation laws for gradient-flow training of conv ResNets and Transformers are characterized for several building blocks, and deep network block laws reduce to laws of isolated blocks.