AutoML for On-Sensor Tiny Machine Learning

Abstract

Sensors with embedded machine learning core (MLC) ena-ble ultra-low-power, low latency, and intelligent inferences at the extreme edge. However, deploying performant machine learning (ML) models within the resource bounds of MLC is challenging. This letter presents an automatic ML framework that trains and deploys ML models on the MLC embedded in the iNEMO inertial measurement units. Given a labeled inertial sensor dataset, the framework automatically selects the best set of features, filters, and window size to apply to the dataset from a search space of supported MLC parameters. The framework then trains a decision tree or random forest within the resource bounds of the MLC and generates the register configuration file to deploy the trained ML model on sensor with MLC capability. For human activity recognition, the framework achieves 93% test accuracy under 1 mW of power, which is 41× lower than an ARM Cortex-M4 implementation. The framework also enables on-sensor fall detection with 95% test accuracy under 0.3 mW of power.