Functional Similarity Metric for Neural Networks: Overcoming Parametric Ambiguity via Activation Region Analysis

Kutomanov Hennadii · 2026

Read Functional Similarity Metric for Neural Networks: Overcoming Parametric Ambiguity via Activation Region Analysis by Kutomanov Hennadii (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-16426.txt.

Introduction

Kutomanov addresses ambiguity in comparing neural networks: different weights can implement the same function, because neurons can be permuted or scaled. Weight-based comparison therefore flickers under small changes. His method instead studies the shape of neuron activation regions, removing scale ambiguity by normalizing weights with layer compensation. Binary signatures of those regions are compared with a fast hashing approximation to the Jaccard index, and neurons are matched by optimal assignment. The metric stays stable under minor weight changes and supports merging, transfer and pruning work.

Catalog record

CREATORKutomanov Hennadii
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE170885
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.16426

Browse the full library · About the Tower of Babel