Distributed Learning for Metaverse over Wireless Networks

Metaverse is envisioned to be a human-centric framework that creates an interface for users to immerse themselves in education, professional training, and entertainment by accessing a virtual world. The quality of immersive experiences (QoIE) naturally comes out as a metric to measure the multi-sensory multimedia (MSMM) communication provided by Metaverse networks, we first propose a human-centric MSMM communication framework and highlight the asymmetric uplink-downlink transmission mechanism by identifying their different responsibilities. This MSMM framework raises the need for advanced communication technologies and more computational resources to support the deployment of AI-enabled Metaverse services. Task-oriented communication (TOC), can enhance conventional data-oriented communication by shifting from data rate maximization to task completion communication, especially deep learning-based TOC (DL-TOC) can build up a joint communication and task completion architecture. The idea of investigating distributed computational resources of end users to perform local learning, and only share model parameters with the central server, known as distributed learning framework, becomes popular, which saves communication resources and provides privacy protection. Then, it is introduced as a beneficial scheme to enable training ML models for both Al-enabled services and the DL-TOC scheme in a distributed manner. Specifically, we propose three distributed learning variants to address the heterogeneity of Metaverse networks from different aspects. Next, a case study is proposed to demonstrate how the proposed distributed learning frameworks can assist attention-aware communication for Metaverse. Finally, we identify the challenges and some promising research directions.