AI Paper Atlas

Research map / Graph learning / GNN architectures

Proper Velocity Neural Networks

Ziheng Chen, Zihan Su, Bernhard Schölkopf, Nicu Sebe

ICLR 2026 · Graph learning › GNN architectures · approach: Graph structure learning · 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 Riemannian PV FC layer instead of tangent-space FC (ablation)PVNN · Airport node classificationaccuracy (%): 86.99 -> 97.93Improved
use Riemannian PV FC layer instead of tangent-space FC (ablation)PVNN · Cora node classificationaccuracy (%): 53.58 -> 52.26Hurt
use GyroBN instead of tangent batch norm (ablation)PVNN · Airport node classificationaccuracy (%): 98.71 -> 99.03Improved
use GyroBN instead of tangent batch norm (ablation)PVNN · Cora node classificationaccuracy (%): 45.36 -> 46.64Improved
lift input features with Exp0 before PVNN (ablation)PVNN · Airport node classificationaccuracy (%): 97.71 -> 97.93No clear effect
lift backbone output with Exp0 in the PV MLR head (ablation)PV MLR · CIFAR-100accuracy (%): 78.20 -> 78.19No clear effect
use tangent activation instead of Euclidean activation in PV (ablation)PVNN · Cora node classificationaccuracy (%): 38.10 -> 52.26Improved
add tangent activation on top of the FC internal nonlinearity (ablation)PVNN · Airport node classificationaccuracy (%): 99.40 -> 99.15No clear effect
use Frechet mean with 10 iterations instead of Euclidean batch statistics (ablation)PV GyroBN · PubMed node classificationaccuracy (%): 69.82 -> 74.34Improved
use Frechet mean with 10 iterations instead of Euclidean batch statistics (cost) (ablation)PV GyroBN · Airport fit timems per epoch: 55.19 -> 105.79Hurt
increase Frechet iterations from 1 to 10 (ablation)PV GyroBN · Airport node classificationaccuracy (%): 88.93 -> 99.03Improved
run Frechet solver to convergence instead of 10 iterations (ablation)PV GyroBN · Airport node classificationaccuracy (%): 99.03 -> 98.46No 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