Rational-activation networks can emulate GELU-type networks with O((log log 1/ε)^3) overhead, while emulating hard rational functions with GELU networks costs Ω(log 1/ε) parameters, an exponential gap.
hub
URL https://link.springer.com/book/10
1 Pith paper cite this work, alongside 1,473 external citations. Polarity classification is still indexing.
1
Pith paper citing it
1,473
external citations · OpenAlex
hub tools
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Rational Neural Networks have Expressivity Advantages
Rational-activation networks can emulate GELU-type networks with O((log log 1/ε)^3) overhead, while emulating hard rational functions with GELU networks costs Ω(log 1/ε) parameters, an exponential gap.