Exploring Local and Cloud-Based Training Use Cases for Embedded Machine Learning

TinyML is a paradigm that makes machine learning (ML) algorithms to run on low-power devices. This involves training an ML model with pre-collected data and then storing it on a microcontroller. In this paper, we explore two application use cases of ML for TinyML devices. First, we present a weather station with a locally-trained snowfall prediction model using TensorFlow Lite for microcontrollers, leveraging on-device temperature and humidity sensors for localized forecasts. Further, a cloud-trained keyword spotting model deployed on an Arduino Nano 33 BLE device demonstrates the use of ML for embedded speech recognition. Our results highlight the need for optimizing parameters and techniques, opening the door to increased precision and efficiency in future embedded ML applications.