MemFLoRA reduces memory needs for CNN training on devices
MemFLoRA: Memory-Floor LoRA for CNN Adaptation at the Edge
Machine LearningArtificial Intelligence
Summary
When smart devices like phones or wearables need to learn from new data on the spot, the biggest hurdle is often memory usage during training, not just the number of settings to adjust. The authors designed MemFLoRA, a way to adapt certain neural networks called CNNs that uses much less memory by changing how training remembers information. They tested it on human activity data and showed it saves almost all the usual memory without losing accuracy.
What this means in practice
- •For mobile app developers: Enable efficient on-device training of CNN models for personalized app features with drastically reduced memory usage.
- •For wearable device engineers: Implement low-memory CNN adaptation methods to handle sensor and environment changes in fitness and health trackers.
Authors
Mehmet Emre Akbulut, Johannes Geier, Ulf Schlichtmann
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.