BROWSE
Detail
Conference Paper
LoRA-Edge: Tensor-Train–Assisted LoRA for Practical CNN Fine-Tuning on Edge Devices
Cited - time in 
Share 
Share
- Authors
- Hyunseok Kwak, Kyeongwon Lee, Jae-Jin Lee, Woojoo Lee
- Issue Date
- 2026-04
- Citation
- Design, Automation & Test in Europe Conference (DATE) 2026, pp.1-7
- Publisher
- European Design and Automation Association
- Language
- English
- Type
- Conference Paper
- Abstract
- On-device fine-tuning of CNNs is essential to withstand domain shift in edge applications such as Human Activity Recognition (HAR), yet full fine-tuning is infeasible under strict memory, compute, and energy budgets. We present LoRA-Edge, a parameter-efficient fine-tuning (PEFT) method that builds on Low-Rank Adaptation (LoRA) with tensor-train assistance. LoRA-Edge (i) applies Tensor-Train Singular Value Decomposition (TT-SVD) to pre-trained convolutional kernels, (ii) selectively updates only the output-side core with zero-initialization to keep the auxiliary path inactive at the start, and (iii) fuses the update back into dense kernels, leaving inference cost unchanged. This design preserves convolutional structure and reduces the number of trainable parameters by up to two orders of magnitude compared to full fine-tuning. Across diverse HAR datasets and CNN backbones, LoRA-Edge achieves accuracy within 4.7% of full fine-tuning while updating at most 1.49% of parameters, consistently outperforming prior parameter-efficient baselines under similar budgets. On a Jetson Orin Nano, TT-SVD initialization and selective-core training yield 1.4-3.8 x faster convergence to target F1. LoRA-Edge thus makes structure-aligned, parameter-efficient on-device CNN adaptation practical for edge platforms.
- KSP Keywords
- Edge devices, Faster convergence, Fine-tuning, HAR datasets, Human activity recognition, Orders of magnitude, Rank adaptation, low rank, singular value decomposition(SVD)