What changes after deployment? A survey on On-device Learning in TinyML

2026-05-29

AI Summary

Machine learning models deployed on microcontroller-level devices are highly susceptible to performance degradation under shifts in data distribution. This work presents the first systematic survey of approximately 70 on-device learning (ODL) studies, unified under a framework categorizing types of distribution shifts. It critically examines how these shifts influence application scenarios, hardware selection, and system architecture. By synthesizing existing literature and establishing a coherent taxonomy, the study focuses on distribution drift modeling, lightweight learning algorithms, and techniques for adapting to resource-constrained hardware. The analysis reveals significant limitations of current approaches in real-world deployments and elucidates how different forms of distribution shift impose critical constraints on ODL system design, thereby bridging the methodological and practical gap between theoretical benchmarks and actual applications.

Abstract

Machine learning models on microcontroller-class devices (TinyML) face a fundamental challenge: post-deployment distribution change undermines static models. On-device learning (ODL) addresses this by running the learning process directly on the device. The existing literature has not characterized how distribution change occurs or how different change types require different solutions. Approximately 70 ODL works are surveyed under one principle: the distribution change regime. The survey analyzes how different types of distribution change influence the applications addressable on-device, the hardware employed, and the structure of the solutions. A persistent gap between methodological benchmarks and real-world deployment scenarios is also identified.

Problem

• TinyML
• On-device Learning
• distribution change
• post-deployment
• microcontroller

Innovation

• On-device Learning
• TinyML
• distribution shift
• microcontroller
• post-deployment adaptation