Logic-based intelligence for batteryless sensors

The emergence of embedded machine learning has enabled the migration of intelligence from the cloud to the edge and to the sensors. To explore the practicalities of wide-spread deployments of these intelligent sensors, we look beyond traditional arithmetic-based neural networks (NNs) to the logic-based learning algorithm called the Tsetlin Machine (TM). TMs have not yet been implemented and explored on general purpose microcontrollers especially that are intermittently powered. In this paper, we argue that their simple architecture makes them a promising candidate for batteryless ML systems. However, in their current form, they are not suitable to be deployed on resource-constrained sensors because of the substantial memory footprint of trained models. To tackle this issue, we propose a lossless compression scheme based on run-length encoding and evaluate against standard TMs for vision and acoustic workloads. We show that our encoding can compress the model by up to 99% without accuracy loss. This translates into lower memory footprint and better energy efficiency (up to 4.9x) compared to the original Tsetlin Machine algorithm, and provides promising trade offs when compared against binary neural networks.