Pierre-Louis Sixdenier, Mark Deutel, Stefan Wildermann, and Prof. Dr.-Ing. Jürgen Teich received the Best Paper Award at this year’s 7th Workshop on IoT, Edge and Mobile for Embedded Machine Learning (ITEM), held in conjunction with ECML-PKDD 2026. Their paper, titled “Early-Exit Forecasting of Deep Neural Networks on Energy-Harvesting Edge Devices,” was recognized with the award.
The prize was presented during the event (October 11, 2026) in Naples.

In this paper, the authors present a two-stage method for selecting the output of an early-exit deep neural network (EE-DNN).
The method is executed periodically on an energy-harvesting microcontroller.

In the first step, a lightweight DNN acting as a forecaster predicts the exit with the highest confidence for a given input.
In the second step, an online energy manager verifies this prediction and dynamically ensures that the resulting energy consumption does not exceed a calculated budget, thereby guaranteeing energy-neutral operation.

The study demonstrates that this combination of forecaster and online energy manager drastically reduces the computational overhead of EE-DNNs and enables energy-neutral operation.
As a result, the runtime of the microcontroller during periodic EE-DNN inference is significantly extended.

⛶

Prof. Dr.-Ing. Jürgen Teich

FAU Research Center Embedded Systems Initiative (FAU ESI)

Head