Adaptive Embedded Machine Learning on IoT Devices: A Review

Embedded Machine Learning (EML) refers to the ability of machine learning models deployed on resource-constrained Internet of Things (IoT) devices to maintain performance over time by adjusting to evolving data distributions and operating conditions. Unlike static models, adaptive EML enables IoT devices to cope with the challenges of dynamic environments, such as concept drift, changing user behavior, and environmental conditions variability. This survey provides a focused overview of recent efforts to enable adaptation in EML, exploring different approaches from system-level strategies like over-the-air (OTA) model update or by directly enabling on-device training of Machine learning (ML) models. We also highlight different learning paradigms, including online learning, continual learning, and meta-learning, that shall improve adaptability in EML. This work synthesizes representative advances in the field and aims to raise awareness of adaptability as a critical but underexplored dimension of EML research. By highlighting key gaps, such as limited time-series adaptation and lack of standardized evaluation, the paper encourages deeper investigation into adaptive mechanisms to support the continued growth of intelligent, resilient, and reliable IoT systems.