Model Fusion via Retrofitting

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Luenam, Phoomraphee, Spanopoulos, Andreas, Sant, Amit, Hofmann, Thomas, Anagnostidis, Sotiris, Singh, Sidak Pal
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913170219597824
author Luenam, Phoomraphee
Spanopoulos, Andreas
Sant, Amit
Hofmann, Thomas
Anagnostidis, Sotiris
Singh, Sidak Pal
author_facet Luenam, Phoomraphee
Spanopoulos, Andreas
Sant, Amit
Hofmann, Thomas
Anagnostidis, Sotiris
Singh, Sidak Pal
contents Model fusion seeks to combine independently trained neural networks into a single model without retraining, but is complicated by representational divergence arising from permutation invariance, random initialization, and heterogeneous training data. Existing methods struggle particularly in zero-shot settings under non-IID data distributions, and are often limited to specific architectures or pairwise fusion. We introduce a neuron-centric family of fusion algorithms that frames fusion as a principled representation-matching problem: intermediate neurons across parent models are grouped into target representations, which the fused model's corresponding sub-networks are then trained to approximate. Unlike prior work, our approach incorporates neuron attribution scores to bias alignment toward salient features, and can be applied to any architecture modularizable as a DAG of levels -- empirically validated on VGGs, ResNets, and ViTs. Experiments across standard benchmarks show consistent improvements over existing fusion methods, with the largest gains in zero-shot and non-IID scenarios. Code is available at https://github.com/AndrewSpano/model-fusion-via-retrofitting.
format Preprint
id arxiv_https___arxiv_org_abs_2507_00037
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Model Fusion via Retrofitting
Luenam, Phoomraphee
Spanopoulos, Andreas
Sant, Amit
Hofmann, Thomas
Anagnostidis, Sotiris
Singh, Sidak Pal
Machine Learning
Artificial Intelligence
I.2.6; I.2.1
Model fusion seeks to combine independently trained neural networks into a single model without retraining, but is complicated by representational divergence arising from permutation invariance, random initialization, and heterogeneous training data. Existing methods struggle particularly in zero-shot settings under non-IID data distributions, and are often limited to specific architectures or pairwise fusion. We introduce a neuron-centric family of fusion algorithms that frames fusion as a principled representation-matching problem: intermediate neurons across parent models are grouped into target representations, which the fused model's corresponding sub-networks are then trained to approximate. Unlike prior work, our approach incorporates neuron attribution scores to bias alignment toward salient features, and can be applied to any architecture modularizable as a DAG of levels -- empirically validated on VGGs, ResNets, and ViTs. Experiments across standard benchmarks show consistent improvements over existing fusion methods, with the largest gains in zero-shot and non-IID scenarios. Code is available at https://github.com/AndrewSpano/model-fusion-via-retrofitting.
title Model Fusion via Retrofitting
topic Machine Learning
Artificial Intelligence
I.2.6; I.2.1
url https://arxiv.org/abs/2507.00037