MemFLoRA: Memory-Floor LoRA for CNN Adaptation at the Edge
The paper says CNN edge adaptation is bottlenecked by saved activations, not just trainable weights.
MemFLoRA is built to keep trainable backward computation from depending on full-width layer inputs. It freezes the down-projection, trains a scale-matched up-projection, and uses eval-mode backbone normalization with activation-minimal backward rules. On three HAR datasets and two CNN backbones, it cuts saved-activation memory by 98.5-98.7% versus full fine-tuning while matching or beating CNN PEFT baselines. Accepted for ASP-DAC 2027. ArXiv · AI/CL/LG's note
MemFLoRA is built to keep trainable backward computation from depending on full-width layer inputs. It freezes the down-projection, trains a scale-matched up-projection, and uses eval-mode backbone normalization with activation-minimal backward rules. On three HAR datasets and two CNN backbones, it cuts saved-activation memory by 98.5-98.7% versus full fine-tuning while matching or beating CNN PEFT baselines. Accepted for ASP-DAC 2027. ArXiv · AI/CL/LG's note
score 4