Toward Scalable Generative AI via Mixture of Experts in Mobile Edge Networks

TL;DR

This paper tackles the challenge of deploying large, resource-intensive generative AI (GAI) on devices with limited resources by integrating Mixture-of-Experts (MoE) with mobile edge networks. It proposes a framework that decomposes GAI tasks, offloads subtasks to edge devices, and uses a Soft Actor-Critic (SAC) based DRL approach to select edge experts while accounting for communication and computation costs. A case study demonstrates that edge-assisted MoE can approach upper-bound QoS and improve content generation quality under constrained resources, validating the practicality of edge-supported GAI. The work highlights a path toward ubiquitous, low-latency, privacy-preserving GAI on mobile devices and outlines future directions in semantic communications, ISAC, and SAGIN-enabled MoE architectures.

Abstract

The advancement of generative artificial intelligence (GAI) has driven revolutionary applications like ChatGPT. The widespread of these applications relies on the mixture of experts (MoE), which contains multiple experts and selectively engages them for each task to lower operation costs while maintaining performance. Despite MoE, GAI faces challenges in resource consumption when deployed on user devices. This paper proposes mobile edge networks supported MoE-based GAI. We first review the MoE from traditional AI and GAI perspectives, including structure, principles, and applications. We then propose a framework that transfers subtasks to devices in mobile edge networks, aiding GAI model operation on user devices. We discuss challenges in this process and introduce a deep reinforcement learning based algorithm to select edge devices for subtask execution. Experimental results will show that our framework not only facilitates GAI's deployment on resource-limited devices but also generates higher-quality content compared to methods without edge network support.

Figures (5)

• Figure 1: The structure of MoE, its advantages, and application examples in DAI and GAI. In the application examples, we illustrate the working principles of MoE through data classification (corresponding to DAI) and text generation (corresponding to GAI).
• Figure 2: The applications of MoE. In this figure, activated links correspond to experts selected by the gating function, utilized for the transmission of tasks and generated content. Inactive links correspond to experts that are not selected and hence do not participate in content generation. In DAI, the MoE can be integrated with models such as multilayer perceptrons and SVMs to enhance the accuracy of classifying images, audio, and even wireless signals, thereby supporting applications such as wireless communication, facial recognition, and more. In GAI, MoE can be combined with Diffusion models and Transformers, so as to improve the quality of generated content, including text, 3D images, audio, and video, thereby enabling applications like chatbots, the metaverse, and among others.
• Figure 3: The operation process of the proposed framework. Upon receiving the prompt, the user device decomposes the task and assesses the computing resources needed for each subtask. Then, considering wireless channel conditions and edge device availability, it selects suitable edge experts and delegates some subtasks to them. Finally, the text generated by the user device is integrated with those from the edge devices to form the final output.
• Figure 4: Reward values versus the number of epoch in DRL.
• Figure 5: Average and final rewards comparison of different methods.