Abstract

Microcontroller Units (MCUs) are ideal platforms for edge applications due to their low cost and energy consumption, and are widely used in various applications, including personalized machine learning tasks, where customized models can enhance the task adaptation. However, existing approaches for local on-device personalization mostly support simple ML architectures or require complex local pre-training/training, leading to high energy consumption and negating the low-energy advantage of MCUs. In this paper, we introduce MicroT, an efficient and low-energy MCU personalization approach. MicroT includes a robust, general, but tiny feature extractor, developed through self-supervised knowledge distillation, which trains a task-specific head to enable independent on-device personalization with minimal energy and computational requirements. MicroT implements an MCU-optimized early-exit inference mechanism called stage-decision to further reduce energy costs. This mechanism allows for user-configurable exit criteria (stage-decision ratio) to adaptively balance energy cost with model performance. We evaluated MicroT using two models, three datasets, and two MCU boards. MicroT outperforms traditional transfer learning (TTL) and two SOTA approaches by 2.12–11.60% across two models and three datasets. Targeting widely used energy-aware edge devices, MicroT's on-device training requires no additional complex operations, halving the energy cost compared to SOTA approaches by up to 2.28× while keeping SRAM usage below 1 MB. During local inference, MicroT reduces energy cost by 14.17% compared to TTL across two boards and two datasets, highlighting its suitability for long-term use on energy-aware resource-constrained MCUs.

What problem does this paper address?

Personalization helps edge models adapt to a user's task or data, but existing on-device methods either support only simple models or impose training costs that erase the energy advantage of MCU deployment. MicroT targets independent local personalization under tight SRAM, compute, and energy budgets.

Main contributions

• A tiny, general feature extractor trained with self-supervised knowledge distillation so an MCU can personalize only a task-specific head.
• A stage-decision early-exit mechanism with a configurable ratio for trading model performance against inference energy.
• An MCU-oriented training path that avoids additional complex local operations and keeps SRAM use below 1 MB.
• Evaluation across two models, three datasets, and two MCU boards.

Key results

• MicroT improved performance by 2.12–11.60% over traditional transfer learning and two compared approaches across the evaluated models and datasets.
• On-device training used up to 2.28× less energy than the compared approaches while keeping SRAM usage below 1 MB.
• Local inference reduced energy cost by 14.17% compared with traditional transfer learning across two boards and two datasets.

Relevant research topics

• TinyML
• On-Device Learning
• Model Personalization
• Microcontrollers
• Self-Supervised Knowledge Distillation
• Early-Exit Inference
• Energy-Efficient Machine Learning
• Edge AI

Preferred citation

@inproceedings{huang2024microt,
  author    = {Yushan Huang and Ranya Aloufi and Xavier Cadet and Yuchen Zhao and Payam Barnaghi and Hamed Haddadi},
  title     = {Low-Energy On-Device Personalization for {MCUs}},
  booktitle = {2024 IEEE/ACM Symposium on Edge Computing (SEC)},
  year      = {2024},
  pages     = {45--58},
  doi       = {10.1109/SEC62691.2024.00012}
}