Enhancing Approximate Message Passing via Diffusion Models Towards On-Device Intelligence
Abstract
In this paper, we introduce a novel denoising-based approximate message passing framework augmented by a noise estimation module realized through a reverse diffusion process. This approach offers significant improvements in training simplicity, recovery quality, and inference efficiency, particularly suitable for deployment at mobile terminals in dynamic environments to facilitate edge computing. Extensive experiments verify the superiority of the proposed method over the state of the art, and demonstrate its potential towards sustainable edge computing and on-device intelligence at scale.