Article’s

Abstract:Tiny machine learning is an emerging approach that enables machine learning models to operate directly on resource-constrained embedded devices. This review aims to examine the principles, technologies, applications, challenges and future directions of machine learning on low-power embedded systems. A structured review of existing research was conducted by examining developments in embedded hardware, software frameworks, model optimisation techniques and application areas. Particular attention was given to quantisation, pruning, knowledge distillation and lightweight neural network architectures, which reduce memory requirements, computational complexity and energy consumption while maintaining useful prediction accuracy. The review also examines applications in healthcare, smart agriculture, environmental monitoring, industrial systems and wearable devices. The findings indicate that local machine learning inference can reduce communication requirements and processing delays while enabling intelligent decision-making on resource-constrained devices. However, limitations in memory, computational capability, energy availability, model accuracy and security continue to present significant challenges. Emerging approaches such as on-device learning, federated learning, energy harvesting and hardware–software co-design provide promising opportunities for addressing these limitations. Overall, the review demonstrates that machine learning on resource-constrained embedded systems has significant potential to transform conventional sensing devices into intelligent, low-power systems capable of real-time local decision-making.