TinyML on Microcontrollers: A Review

Abstract

Tiny Machine Learning (TinyML) has emerged as a potential paradigm for deploying and training machine learning models on resource-constrained systems like microcontroller units (MCUs) to enable real-time and low-power intelligent systems at the edge. This paper presents a comprehensive review on the current landscape of TinyML, focusing on architectural considerations of MCUs and machine learning models, model optimisation techniques, deployment techniques, and potential application-specific implementations of TinyML. Among the proposed models and frameworks are MCUNet, EfficientNet, TinyFL, TinyOL, TinyOPs and many more, which explore approaches to enhancing the efficacy, model training, and optimisation techniques for deploying deep neural networks on MCUs. We explore the different trade-offs in training and compression methodologies, techniques used for optimising inference for efficient execution of models and focus on exploring the feasibility of on-device training on MCUs so that the model can adapt to real-time sensing and collation of data. Furthermore, we examine the possibilities of real-world applications spanning healthcare, industrial automation and smart environments while focusing on the potential and limitations of TinyML in different domains. By reviewing the current situation of the various aforementioned factors in TinyML, this paper aims to provide a foundational insight into the future directions of study in this field.

DOI: https://doi.org/10.24321/2456.1428.202603