arXiv cs.AI· Mehmet Emre Akbulut, Johannes Geier, Ulf Schlichtmann·· 10 天前AI 评分22
MemFLoRA：面向边缘端 CNN 适配的内存下限 LoRA
MemFLoRA: Memory-Floor LoRA for CNN Adaptation at the Edge
AI 导读
MemFLoRA 是一种面向 CNN 边缘适配的低秩适配器，以“内存优先”为设计原则，冻结下投影、训练尺度匹配的上投影，并将反向传播所需保存的状态压缩到低秩分支。在三个 HAR 数据集和两个 CNN 骨干上，其保存激活内存较全量微调降低 98.5-98.7%，峰值训练状态内存降低 94.9-97.3%，同时匹配或超过 CNN PEFT 基线。该工作已被 ASP-DAC 2027 接收。
正文 · 原文
Abstract:On-device learning is necessary when the model encounters user-,sensor-, or environment-specific shifts after deployment. Although parameter-efficient fine-tuning (PEFT) methods, particularly Low-Rank Adaptation (LoRA) variants, enable efficient adaptation at the edge, the limiting resource for Convolutional Neural Network (CNN) adaptation is often not the number of trainable parameters but the activation state that must be retained until the backward pass. This paper introduces Memory-Floor LoRA (MemFLoRA), a low-rank CNN adapter built around a memory-first design principle rather than a direct application of transformer-oriented LoRA. Instead of merely reducing trainable weights, we define an activation-memory-floor criterion: trainable backward computations must not depend on full-width layer inputs. The resulting adapter freezes the down-projection, trains a scale-matched up-projection, and combines eval-mode backbone normalization with activation-minimal backward rules, reducing saved state to the low-rank branch. Evaluated on three Human Activity Recognition (HAR) datasets and two CNN backbones under subject, body-location, and sensor-placement shifts, MemFLoRA reduces saved-activation memory by 98.5-98.7% and peak training-state memory by 94.9-97.3% relative to full fine-tuning, while matching or exceeding CNN PEFT baselines.
| Comments: | Accepted at the 32nd Asia and South Pacific Design Automation Conference (ASP-DAC 2027), January 25-28, 2027, Tokyo, Japan. Code: this https URL |
| Subjects: | Machine Learning (cs.LG); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.08669 [cs.LG] |
| (or arXiv:2610.08669v1 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2610.08669 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Mehmet Emre Akbulut [view email]
[v1]
Tue, 6 Oct 2026 16:51:20 UTC (3,934 KB)
来源：arXiv cs.AI · arxiv.org