What changes after deployment? A survey on On-device Learning in TinyML
This work is significant for TinyML researchers and practitioners, as it characterizes how distribution changes occur and how different types necessitate different ODL solutions, highlighting a gap between benchmarks and real-world scenarios.
This survey paper investigates the challenges of post-deployment distribution shifts in TinyML, which static models cannot address. It analyzes approximately 70 on-device learning (ODL) works, categorizing them by the type of distribution change they tackle and its impact on applications, hardware, and solution structures.
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.