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

Purpose and goal

One of the key visions of 6G is to enable connectivity for a massive number of IoT devices, far beyond the capabilities of current networks such as 4G and 5G. However, today’s IoT devices predominantly rely on models trained in the cloud, which makes it difficult to adapt them to the specific characteristics and dynamics of their deployment environments. The objective of this work is to enable on-device training, thereby allowing models to be continuously tailored to their local context.

Expected effects and result

The expected result is to demonstrate the feasibility of online learning on Ambient IoT devices by designing and implementing an online learning system that enables model training on Ambient IoT devices. On-device training allows (1) training models via local data without sharing data, thus enabling privacy-preserving computation by design, (2) model personalization and environment adaptation, and (3)deploying accurate models in any location without stable internet connectivity.

Planned approach and implementation

After a literature review covering the theoretical foundations and implementation details of Mondrian Forests, the next step is to demonstrate the feasibility of the implementation that does not require storing the entire dataset used for pre-training. This implementation will likely need to be optimized for the resource constraints of IoT devices, such as memory usage, and power consumption. Finally, the proposed approach will be evaluated, and the thesis will be documented and defended.