TinyMLOps for real-time ultra-low power MCUs applied to frame-based event classification

Abstract

TinyML applications such as speech recognition, motion detection, or anomaly detection are attracting many industries and researchers thanks to their innovative and cost-effective potential. Since tinyMLOps is at an even earlier stage than MLOps, the best practices and tools of tinyML are yet to be found to deliver seamless production-ready applications. TinyMLOps has common challenges with MLOps, but it differs from it because of its hard footprint constraints. In this work, we analyze the steps of successful tinyMLOps with a highlight on challenges and solutions in the case of real-time frame-based event classification on low-power devices. We also report a comparative result of our tinyMLOps solution against tf.lite and NNoM.