Model Training for Tiny Machine Learning

Abstract

The integration of machine learning (ML) and embedded systems is referred to as tiny machine learning (TinyML), which aims to achieve intelligent functions on ultra-low power devices. Such devices are typically bound by computation and memory; however, with the help of sensors, they can physically monitor the surrounding environment and act based on the decisions made through the minimum ML algorithms (TinyML). There is a close relation between the performance of the deployment platform and the TinyML framework, as they work together towards achieving power and performance efficiency. The primary objective of implementing TinyML on microcontrollers is to allow the device to perform ML computations independently of cloud services, at the edge. This enhances the speed of operation, minimizes power consumption, and protects confidential information. The foundation of TinyML promotes the development of TinyML technology internationally, wherein ML is applied toward solving global issues, accompanied by energy-efficient powering solutions. To optimize the functionality of advanced devices in limited computing environments, techniques such as quantization and pruning are utilized to reduce the size and augment the speed of neural network inference. The idea of TinyML does not include model training on microcontrollers; rather, it suggests uploading pre-trained models that are stored in an optimized format on other systems.