AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Exploiting weight-space symmetries for approximating curvature

Artem Artemev, Rui Xia, Benjamin M. Boyd, Youjing Yu, Felix Dangel, Guillaume Hennequin, Alberto Bernacchia

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Symmetry and equivariance in loss landscapes · 0 citations (OpenAlex)

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 block-diagonal restriction of H*_g (ablation)Symo Hessian approximation · Teacher-student MLP (D=19,850), random-direction secant cosine similaritycosine similarityNo clear effect
use identity weight covariance (H*_g) instead of inverting it (H*_PD) (ablation)Symo Hessian approximation · Gradient-direction secant testcosine similarityImproved
use Hessian at orbit-averaged parameters H* alone (ablation)Symo Hessian approximation · Teacher-student MLP (D=19,850), random-direction secant cosine similaritycosine similarity: between 0.1 and 0.2 -> near zeroHurt
use Symo instead of AdamMNIST autoencoder and nanoGPT · MNIST autoencoder / nanoGPTtraining curvesImproved
use largest symmetry group (all layers permuted) in Symo (ablation)Symo · Deep MLP autoencodertraining curvesImproved
use smaller group with odd-layer permutation and block-diagonal Hessian (ablation)Symo · Deep MLP autoencodertraining curvesNo clear effect

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