Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Low-Energy Adaptive Personalization for Resource-Constrained Devices

Yushan Huang, Josh Millar, Yuxuan Long, Yuchen Zhao, Hamed Haddadi

TL;DR

This work introduces Target Block Fine-Tuning (TBFT), a low-energy on-device personalization framework that selects model blocks to fine-tune based on drift type: input-level, feature-level, or output-level. By freezing non-selected blocks and only updating the front, middle, or rear blocks accordingly, TBFT achieves substantial accuracy improvements while reducing energy usage compared with full fine-tuning. Evaluations on ResNet-26 across CIFAR-10-C, Living17, and CIFAR-Flip datasets on a Raspberry Pi demonstrate an average accuracy gain of about $15.30\%$ and energy savings around $41.57\%$, validating the method’s effectiveness for resource-constrained environments. The approach balances personalization performance and energy efficiency, with discussions on drift-detection enhancements, unsupervised extensions, and multidimensional drift scenarios for future work.

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 the $Block Avg$, 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.

Towards Low-Energy Adaptive Personalization for Resource-Constrained Devices

TL;DR

This work introduces Target Block Fine-Tuning (TBFT), a low-energy on-device personalization framework that selects model blocks to fine-tune based on drift type: input-level, feature-level, or output-level. By freezing non-selected blocks and only updating the front, middle, or rear blocks accordingly, TBFT achieves substantial accuracy improvements while reducing energy usage compared with full fine-tuning. Evaluations on ResNet-26 across CIFAR-10-C, Living17, and CIFAR-Flip datasets on a Raspberry Pi demonstrate an average accuracy gain of about and energy savings around , validating the method’s effectiveness for resource-constrained environments. The approach balances personalization performance and energy efficiency, with discussions on drift-detection enhancements, unsupervised extensions, and multidimensional drift scenarios for future work.

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 the , 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.
Paper Structure (14 sections, 2 equations, 3 figures, 4 tables)

This paper contains 14 sections, 2 equations, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Adaptive Personalization
  4. Implementation & Evaluation Setup
  5. Model and Datasets
  6. Prototype and Performance Metrics
  7. Preliminary Evaluation Results
  8. Motivation Experiments
  9. Overall Performance
  10. Single-Batch & Multi-Batch Training
  11. Model Performance
  12. System Cost
  13. Discussion and Future Work
  14. Conclusion

Figures (3)

  • Figure 1: System Overview.
  • Figure 2: The ResNet-26 model used for all experiments.
  • Figure 3: The accuracy results of TBFT on three datasets and with different training sizes. The horizontal solid line shows the accuracy of $Block Avg$, and the horizontal dashed line the accuracy of full model fine-tuning. The highlighted values are the best accuracy across block-based fine-tuning.