.png){kind=icon}
.png){kind=icon}
Return
Resource-aware collaborative learning with improved generalization on edge devices
.png){kind=icon}
PRE
AI
DOI：10.1016/j.sysarc.2026.103935.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
While deploying LLMs on edge devices enables low-latency and privacy-preserving intelligent services, collaborative learning across heterogeneous devices faces two key challenges: heterogeneous update drift and sharpness-induced overfitting. To address these, we propose RASP, a resource-aware framework that optimizes perturbed submodels via a minimax objective. RASP prunes the global model into adaptive sizes to fit local resource while suppressing sharp local minima to enhance generalization. Theoretically, we prove an asymptotically optimal convergence rate of O(1/QTC∗) for non-convex objectives, where C∗ is the minimum covering number. We also derive a tighter generalization bound that characterizes the impact of the perturbation magnitude and layer-wise parameter retention rates. Extensive experiments show that RASP consistently improves global performance and stability, outperforming state-of-the-art baselines in heterogeneous edge environments.
Journal
IF:
4.1
Papers:
3.0K
Citations:
4.2K