A closed-form initialization for SIREN networks based on pre-activation fixed points and Jacobian variance sequences improves gradient scaling, training dynamics via NTK, and generalization on reconstruction tasks over the original scheme.
Since − 1 e <− c2w 3 e−c2w/3−2c2 b <0, the properties of the principal branch W0 imply |f ′(σ 2a )|< 1
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
A new initialisation to Control Gradients in Sinusoidal Neural network
A closed-form initialization for SIREN networks based on pre-activation fixed points and Jacobian variance sequences improves gradient scaling, training dynamics via NTK, and generalization on reconstruction tasks over the original scheme.