Resource-aware online learning for resource-constrained 6G Ambient IoT devices

Important results from the project

The main project goal was achieved by demonstrating the feasibility of Mondrian Forests, an online tree-ensemble method designed for incremental learning from streaming data, on resource-constrained Ambient IoT devices using a bare-metal STM32 MCU. The project also introduced memory optimizations, including quantization and a flash-assisted technique for reducing RAM usage. However, results indicate that robust adaptation to concept drift requires further work.

Expected long term effects

The project is expected to support autonomous, privacy-preserving, and adaptive Ambient IoT systems. By demonstrating the feasibility of Mondrian Forests for online learning directly on a bare-metal STM32 MCU, the work points toward systems that can learn locally without reliance on cloud infrastructure. It also provides a basis for further research on robust on-device learning, particularly on improving the ability of Mondrian Forests to adapt to changing environments and concept drift.

Approach and implementation

The project followed the planned process, from literature review to completed thesis and defense. The work began with a study of Mondrian Forests, online learning, and IoT constraints. This was followed by implementation, optimization for MCU resources, and experimental evaluation. The results were documented in the thesis, which was presented and defended according to plan. No major external factors affected the work, and the collaboration functioned well.