Embedded Machine Learning for Microcontroller-Class Edge Devices: Data, Feature, Evaluation, and Deployment Pipelines

The paper presents a comprehensive workflow for implementing embedded machine learning on microcontroller-class edge devices, focusing on data acquisition, preprocessing, model execution, and deployment under constraints of memory, energy, and latency. It details the engineering decisions necessary for effective model design, including feature extraction techniques for inertial motion recognition and keyword spotting using a compact one-dimensional convolutional network. This work is significant for practitioners as it provides practical design rules and methodologies for optimizing on-device inference, which are critical for developing efficient edge AI applications.