.png){kind=icon}
.png){kind=icon}
Return
Edge-PRUNE: A Dataflow-Based Framework for Distributed Signal Processing and Machine Learning
.png){kind=icon}
OA
AI
DOI：10.1109/TSP.2025.3598453.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Distributed sensing through video, audio, radar and other sensors is strongly growing with application areas such as smart homes and Internet of Things. The concept of edge computing proposes shifting signal and data analysis from centralized servers close to the sensors, providing reduction in data communication bandwidth requirements and centralized server computation load as well as improving data privacy. Previous works in the domain of edge computing have paid little attention to formal modeling of computing across devices. This work proposes the VR-PRUNE-E model of computation that is based on the well-known dataflow abstraction. Within VR-PRUNE-E, a specific type of resilient network graph is introduced, which allows the distributed system to continue its operation after the failure of any single node or connection. Besides the formal model, the manuscript introduces the Edge-PRUNE software framework that supports the proposed dataflow abstraction, as well as concrete experimental results on real edge computing scenarios. The explored setups cover networks with up to 128 endpoint nodes and two servers. Application examples cover popular machine learning applications of image classification, object detection and radar signal processing, built on CNN and transformer architectures, extended with redundant system configurations that provide fault tolerance. The proposed work is also benchmarked in terms of processing time and shown to outperform previous work by 34% in computation efficiency.
Keywords:
Dataflow computing
signal processing
machine learning
edge computing
fault tolerance
Journal
IF:
13.7
Papers:
1.0W
Citations:
8.4W
View more organizations