Topological Information Residency in Artificial Neural Networks

Edge-Centric Computation and Empirical Robustness of Dodecaflake Architectures

Author :
Guillaume Desvaux · HOPE 'N MIND SASU
Year :
2026
Reading time :
18 min

Abstract

Experimental evidence that information in trained neural networks resides topologically within edges (weights) rather than nodes (neurons). Introduces Hexaflake and Dodecaflake architectures: 75% of propagation facets can be ablated with zero performance loss, demonstrating unprecedented topological protection.

Keywords

  • Mechanistic interpretability
  • Memory & information
  • Mathematical complexity
  • Falsifiability
  • Neural networks & LLMs
  • Edge-centric architectures
  • Robustness & topological resilience
  • Complex Systems & Metrology
  • Falsifiability & epistemology
Read the interactive articleView on Zenodo (PDF)