Abstract

The personalization of machine learning (ML) models to address data drift is a significant challenge in the context of Internet of Things (IoT) applications. Presently, most approaches focus on fine-tuning either the full base model or its last few layers to adapt to new data, while often neglecting energy costs. However, various types of data drift exist, and fine-tuning the full base model or the last few layers may not result in optimal performance in certain scenarios. We propose Target Block Fine-Tuning (TBFT), a low-energy adaptive personalization framework designed for resource-constrained devices. We categorize data drift and personalization into three types: input-level, feature-level, and output-level. For each type, we fine-tune different blocks of the model to achieve optimal performance with reduced energy costs. Specifically, input-, feature-, and output-level correspond to fine-tuning the front, middle, and rear blocks of the model. We evaluate TBFT on a ResNet model, three datasets, three different training sizes, and a Raspberry Pi. Compared with BlockAvg, where each block is fine-tuned individually and their performance improvements are averaged, TBFT exhibits an improvement in model accuracy by an average of 15.30% whilst saving 41.57% energy consumption on average compared with full fine-tuning.

What problem does this paper address?

Distribution shifts can occur at the input, feature, or output level, yet common personalization methods always fine-tune the full network or its final layers. That fixed choice can waste energy and adapt the wrong part of a model on a resource-constrained edge device.

Main contributions

• A three-part taxonomy linking input-, feature-, and output-level drift to front, middle, and rear model blocks.
• Target Block Fine-Tuning, which chooses the model region to update according to the detected personalization scenario.
• An evaluation spanning a ResNet, three datasets, three training-set sizes, and Raspberry Pi energy measurements.

Key results

• TBFT improved accuracy by 15.30% on average relative to the BlockAvg comparison.
• TBFT saved 41.57% energy on average relative to full-model fine-tuning.

Relevant research topics

• On-Device Personalization
• Data Drift
• Targeted Fine-Tuning
• Resource-Constrained Devices
• Edge AI
• Energy-Efficient Learning
• TinyML

Preferred citation

@inproceedings{huang2024tbft,
  author    = {Yushan Huang and Josh Millar and Yuxuan Long and Yuchen Zhao and Hamed Haddadi},
  title     = {Towards Low-Energy Adaptive Personalization for Resource-Constrained Devices},
  booktitle = {Proceedings of the 4th Workshop on Machine Learning and Systems},
  year      = {2024},
  pages     = {73--80},
  doi       = {10.1145/3642970.3655826}
}