WiP Paper

eODL: Energy-Aware On-Device Training on Microcontrollers

Conference Paper (2026)
Author(s)

Hao Liu (Student TU Delft)

Shashwath Suresh (Student TU Delft)

Marco Zuniga (TU Delft - Electrical Engineering, Mathematics and Computer Science)

Research Group
Networked Systems
DOI related publication
https://doi.org/10.3217/j1ep-fx24 Final published version
More Info
expand_more
Publication Year
2026
Language
English
Research Group
Networked Systems
Journal title
International Conference on Embedded Wireless Systems and Networks
Volume number
2026
Pages (from-to)
1-6
Event
23rd International Conference on Embedded Wireless Systems and Networks, EWSN 2026 (2026-09-16 - 2026-09-18), Dresden, Germany
Page Views
17
Reuse Rights

Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.

Abstract

Deep learning models can be trained on microcontrollers (MCUs) for privacy-sensitive applications, but such training is severely constrained by tiny SRAM capacity and limited battery energy. Existing on-device learning methods optimise memory usage by selectively freezing layers so that training fits within SRAM budgets, yet none explicitly account for training energy. A strategy that is feasible in memory may therefore still consume significantly more energy than necessary, shortening battery life on energyconstrained devices. We present eODL, an energy-aware on-device learning framework. eODL exploits two observations: (1) counts of multiplyaccumulate operations (MACs) predict training energy with high fidelity (R2 = 0.993), so energy can be estimated; and (2) energyaware layer-update selection can be solved efficiently with tabular dynamic programming. Using these, eODL selects which layers to update under a hard SRAM budget while balancing energy against accuracy. We implement eODL end to end on STM32 MCUs and evaluate it on MCUNetV2 fine-tuned on the Flowers-102 dataset. Under a 50 kB training-memory budget, eODL matches state-of-the-art accuracy while reducing measured on-device energy by 22.6% per epoch.