TinyML On-Device Learning Survey Categorizes Distribution Change
A new survey examines approximately 70 works on on-device learning (ODL) for TinyML, focusing on how post-deployment distribution change affects machine learning models on microcontroller-class devices. The survey categorizes distribution change types and analyzes their impact on applications, hardware, and solution structures. It identifies a persistent gap between methodological benchmarks and real-world deployment scenarios.
Key facts
- Machine learning models on microcontroller-class devices (TinyML) face post-deployment distribution change.
- On-device learning (ODL) runs the learning process directly on the device.
- Approximately 70 ODL works are surveyed.
- The survey is organized under the principle of distribution change regime.
- Different types of distribution change require different solutions.
- The survey analyzes how distribution change influences applications, hardware, and solution structures.
- A persistent gap between methodological benchmarks and real-world deployment is identified.
- The survey is from arXiv, submission history not specified.
Entities
Institutions
- arXiv