An energy functional is constructed so its stationary point is softmax attention, then a regularizer is added by hand to force the same point for nonlinear variants, with no valid derivation or experiments.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
stat.ML 1years
2025 1verdicts
REJECT 1representative citing papers
citing papers explorer
-
A Framework for Non-Linear Attention via Modern Hopfield Networks
An energy functional is constructed so its stationary point is softmax attention, then a regularizer is added by hand to force the same point for nonlinear variants, with no valid derivation or experiments.