An Embedded On-Device Training Framework for Neural Networks

Michael Benjamin Kargl; David Josef Muttenthaler; Florian Eibensteiner; Christoph Dalpiaz; Phillip Michael Christian Fred Kastner; Josef Langer

An Embedded On-Device Training Framework for Neural Networks

Michael Benjamin Kargl
, David Josef Muttenthaler
, Florian Eibensteiner
, Christoph Dalpiaz
, Phillip Michael Christian Fred Kastner
, Josef Langer

Research Center Hagenberg

Publikation: Konferenzbeitrag › Papier › Begutachtung

Abstract

We present a portable framework that enables inference and training of neural networks directly on embedded devices. It consists of a PyTorch model parser that exports networks to a binary format and a C++ runtime capable of executing inference and training on this binary format. A run of a minimal CNN on MNIST achieved numerical parity with PyTorch in inference and demonstrated the feasibility of our framework for on-device fine-tuning in non-time-constrained scenarios on an STM32F469 microcontroller.

Originalsprache	Englisch
Publikationsstatus	Angenommen/Im Druck - 2026
Veranstaltung	20th International Conference on Computer Aided Systems Theory - Museo Elder de la Ciencia y la Tecnología, Las Palmas de Gran Canaria, Spanien Dauer: 23 Feb. 2026 → 27 Feb. 2026 https://eurocast2026.fulp.es/

Konferenz

Konferenz	20th International Conference on Computer Aided Systems Theory
Kurztitel	Eurocast 2026
Land/Gebiet	Spanien
Ort	Las Palmas de Gran Canaria
Zeitraum	23.02.2026 → 27.02.2026
Internetadresse	https://eurocast2026.fulp.es/

Zitieren

@conference{897da04848744c97a75ad1731e316758,

title = "An Embedded On-Device Training Framework for Neural Networks",

abstract = "We present a portable framework that enables inference and training of neural networks directly on embedded devices. It consists of a PyTorch model parser that exports networks to a binary format and a C++ runtime capable of executing inference and training on this binary format. A run of a minimal CNN on MNIST achieved numerical parity with PyTorch in inference and demonstrated the feasibility of our framework for on-device fine-tuning in non-time-constrained scenarios on an STM32F469 microcontroller.",

keywords = "embedded ML, on-device training, tinyML",

author = "Kargl, \{Michael Benjamin\} and Muttenthaler, \{David Josef\} and Florian Eibensteiner and Christoph Dalpiaz and Kastner, \{Phillip Michael Christian Fred\} and Josef Langer",

year = "2026",

language = "English",

note = "20th International Conference on Computer Aided Systems Theory, Eurocast 2026 ; Conference date: 23-02-2026 Through 27-02-2026",

url = "https://eurocast2026.fulp.es/",

}

TY - CONF

T1 - An Embedded On-Device Training Framework for Neural Networks

AU - Kargl, Michael Benjamin

AU - Muttenthaler, David Josef

AU - Eibensteiner, Florian

AU - Dalpiaz, Christoph

AU - Kastner, Phillip Michael Christian Fred

AU - Langer, Josef

PY - 2026

Y1 - 2026

N2 - We present a portable framework that enables inference and training of neural networks directly on embedded devices. It consists of a PyTorch model parser that exports networks to a binary format and a C++ runtime capable of executing inference and training on this binary format. A run of a minimal CNN on MNIST achieved numerical parity with PyTorch in inference and demonstrated the feasibility of our framework for on-device fine-tuning in non-time-constrained scenarios on an STM32F469 microcontroller.

AB - We present a portable framework that enables inference and training of neural networks directly on embedded devices. It consists of a PyTorch model parser that exports networks to a binary format and a C++ runtime capable of executing inference and training on this binary format. A run of a minimal CNN on MNIST achieved numerical parity with PyTorch in inference and demonstrated the feasibility of our framework for on-device fine-tuning in non-time-constrained scenarios on an STM32F469 microcontroller.

KW - embedded ML

KW - on-device training

KW - tinyML

M3 - Paper

T2 - 20th International Conference on Computer Aided Systems Theory

Y2 - 23 February 2026 through 27 February 2026

ER -

An Embedded On-Device Training Framework for Neural Networks

Abstract

Konferenz

Fingerprint

Zitieren