AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

A New Initialization to Control Gradients in Sinusoidal Neural Networks

Andrea Combette, Nelly Pustelnik, Antoine Venaille

ICLR 2026 · Architectures and efficiency › RNNs and normalization · approach: Theory of depth and width

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use proposed SIREN init (c_w=sqrt3, c_b=0, sigma_a=0) instead of original Sitzmann init (ablation)SIREN · image fitting and function approximation across depthsreconstruction noise / generalizationImproved
use sigma_a=0 instead of sigma_a=1 along sigma_g=1 curve (ablation)SIREN proposed init · deep image fittingoutput noise / spurious high frequenciesImproved
use proposed init instead of original SIREN init at large depth (ablation)SIREN · NTK trace and input gradient vs depthgradient norm / NTK eigenvalues: original grows exponentially (sigma_g about sqrt(1.2)); proposed sigma_g = 1 constant gradientsImproved
use PyTorch default init instead of proposed init (ablation)SIREN · NTK trace and input gradient vs depthgradient norm: sigma_g = sqrt(1/3)Hurt
use proposed init (sigma_a=0) for Fourier spectrum control (ablation)SIREN · 1D Fourier spectra, depths 4-32spectral energy beyond w0Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy