Virtual neural networks: hundreds of souls in a body

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F61988987%3A17610%2F25%3AA2602FBZ" target="_blank" >RIV/61988987:17610/25:A2602FBZ - isvavai.cz</a>

Výsledek na webu

<a href="https://link.springer.com/article/10.1007/s00521-025-11180-y" target="_blank" >https://link.springer.com/article/10.1007/s00521-025-11180-y</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.1007/s00521-025-11180-y" target="_blank" >10.1007/s00521-025-11180-y</a>

Jazyk výsledku

angličtina

Název v původním jazyce

Virtual neural networks: hundreds of souls in a body

Popis výsledku v původním jazyce

We propose a novel paradigm called virtual neural networks where the number of trainable parameters is fixed and the scalability is made on the computation cost level only. The paradigm is an abstract structure that can be implemented using an arbitrary standard convolutional neural network. It combines siamese neural networks and a deep ensemble approach by creating many virtual models that share combined weights given by a few physical models. The ensemble consists of up to hundreds of virtual models that are trained concurrently. Moreover, all virtual networks share the same input, and their tangled structure creates a kind of inner augmentation that elevates the robustness of the whole ensemble. The accuracy of the ensemble increases with the number of virtual networks, while the capacity remains the same. We demonstrate that virtual neural networks outperform models with larger capacity, standard deep ensembles, and modern techniques such as SWA and Masksembles. Moreover, the best single model from our trained ensemble produces better results than other single-trained models, even with more parameters. The code is available online at gitlab.com/EnginCZ/virtual-models-public.

Název v anglickém jazyce

Virtual neural networks: hundreds of souls in a body

Popis výsledku anglicky

We propose a novel paradigm called virtual neural networks where the number of trainable parameters is fixed and the scalability is made on the computation cost level only. The paradigm is an abstract structure that can be implemented using an arbitrary standard convolutional neural network. It combines siamese neural networks and a deep ensemble approach by creating many virtual models that share combined weights given by a few physical models. The ensemble consists of up to hundreds of virtual models that are trained concurrently. Moreover, all virtual networks share the same input, and their tangled structure creates a kind of inner augmentation that elevates the robustness of the whole ensemble. The accuracy of the ensemble increases with the number of virtual networks, while the capacity remains the same. We demonstrate that virtual neural networks outperform models with larger capacity, standard deep ensembles, and modern techniques such as SWA and Masksembles. Moreover, the best single model from our trained ensemble produces better results than other single-trained models, even with more parameters. The code is available online at gitlab.com/EnginCZ/virtual-models-public.

Druh

J_SC - Článek v periodiku v databázi SCOPUS

CEP obor

—

OECD FORD obor

10102 - Applied mathematics

Projekt

<a href="/cs/project/EH22_008%2F0004583" target="_blank" >EH22_008/0004583: Excelentní výzkum v oblasti digitálních technologií a wellbeingu</a><br>

Návaznosti

P - Projekt vyzkumu a vyvoje financovany z verejnych zdroju (s odkazem do CEP)

Rok uplatnění

2025

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Název periodika

Neural Computing and Applications

ISSN

0941-0643

e-ISSN

1433-3058

Svazek periodika

—

Číslo periodika v rámci svazku

19

Stát vydavatele periodika

GB - Spojené království Velké Británie a Severního Irska

Počet stran výsledku

19

Strana od-do

14279-14297

Kód UT WoS článku

—

EID výsledku v databázi Scopus

2-s2.0-105004902717