Saved in:
Bibliographic Details
Main Authors: Jeong, Kyungwon, Paeng, Won-Gi, Suh, Honggyo
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2605.04971
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915984070148096
author Jeong, Kyungwon
Paeng, Won-Gi
Suh, Honggyo
author_facet Jeong, Kyungwon
Paeng, Won-Gi
Suh, Honggyo
contents Weight matrices in deep networks exhibit geometric continuity -- principal singular vectors of adjacent layers point in similar directions. While this property has been widely observed, its origin remains unexplained. Through experiments on toy MLPs and small transformers, we identify two mechanisms: residual connections create cross-layer gradient coherence that aligns weight updates across layers, and symmetry-breaking nonlinearities constrain all layers to a shared coordinate frame, preventing the rotation drift that would otherwise destabilize weight structure. Crucially, a nonlinear but rotation-preserving activation fails to retain continuity, isolating symmetry breaking -- not nonlinearity itself -- as the active ingredient. Activation and normalization play distinct roles: activation concentrates continuity in the leading singular direction, while normalization distributes it across multiple directions. In transformers, continuity is projection-specific: Q, K, Gate, and Up (which read from the residual stream) develop input-space ($\mathbf{v}_1$) continuity; O and Down (which write to it) develop output-space ($\mathbf{u}_1$) continuity; V alone, lacking an adjacent nonlinearity, develops only low continuity.
format Preprint
id arxiv_https___arxiv_org_abs_2605_04971
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Why Geometric Continuity Emerges in Deep Neural Networks: Residual Connections and Rotational Symmetry Breaking
Jeong, Kyungwon
Paeng, Won-Gi
Suh, Honggyo
Machine Learning
Artificial Intelligence
Computation and Language
68T07
I.2.6; I.2.7
Weight matrices in deep networks exhibit geometric continuity -- principal singular vectors of adjacent layers point in similar directions. While this property has been widely observed, its origin remains unexplained. Through experiments on toy MLPs and small transformers, we identify two mechanisms: residual connections create cross-layer gradient coherence that aligns weight updates across layers, and symmetry-breaking nonlinearities constrain all layers to a shared coordinate frame, preventing the rotation drift that would otherwise destabilize weight structure. Crucially, a nonlinear but rotation-preserving activation fails to retain continuity, isolating symmetry breaking -- not nonlinearity itself -- as the active ingredient. Activation and normalization play distinct roles: activation concentrates continuity in the leading singular direction, while normalization distributes it across multiple directions. In transformers, continuity is projection-specific: Q, K, Gate, and Up (which read from the residual stream) develop input-space ($\mathbf{v}_1$) continuity; O and Down (which write to it) develop output-space ($\mathbf{u}_1$) continuity; V alone, lacking an adjacent nonlinearity, develops only low continuity.
title Why Geometric Continuity Emerges in Deep Neural Networks: Residual Connections and Rotational Symmetry Breaking
topic Machine Learning
Artificial Intelligence
Computation and Language
68T07
I.2.6; I.2.7
url https://arxiv.org/abs/2605.04971