Joint Semantic Information Gathering and Broadcasting for Multi-task Edge Processing

In this paper, we propose an edge computing system where an edge server (ES) gathers multi-media sensing data from collaborative sensing devices (SDs) in upstream and broadcasts the sensing information to downstream user devices that perform different tasks, e.g., image segmentation and saliency detection. In particular, we leverage the wisdom of semantic communication that applies joint source-channel coding (JSCC) techniques to extract and transmit only the task-relevant information contained in the sensing data to reduce the communication and computation workload. To efficiently train the upstream and downstream encoders/decoders involved in the system, instead of performing end-to-end training, we propose a task-oriented training strategy that decouples the training into two separate steps without compromising the performance. We conduct experiments using open-source multi-task datasets under varying channel signal-to-noise power ratios (SNRs). The simulation results show that, compared to the representative benchmark approaches, the proposed method can effectively reduce $\mathbf{9 2. 6 \%} / 75 \%$ uplink/downlink data workload and elegantly balance the performance of different tasks.

I. Introduction

A complete wireless edge computing system typically includes edge sensing devices (SDs), an edge server (ES), and users (Fig. 1). Upstream edge SDs collect sensing data and transmit it to the ES through uplink channels. The ES integrates data from the SDs and disseminates to downstream users. Due to advancements in communication and internet of things (IoT) technologies, the number and variety of edge devices have increased dramatically on the upstream side. Additionally, the transmitted data is of higher bandwidth, including high-resolution videos and 3D images. Directly transmitting raw data would overload the upstream channel [1]. On the downstream side, the growing number of users puts an increased workload on the ES with limited computing power [2], [3]. Therefore, for wireless upstream and downstream edge computing systems like the one in Fig. 1, a joint design method is crucial to optimize information gathering and dissemination. This method should ensure accurate downstream task execution while minimizing communication resource consumption during data transmission.