JEPA Architecture Explored for AI-Native 6G Networks

Sixth-generation (6G) networks are evolving towards an AI-native operational model, integrating learning modules throughout the radio access network, edge, and core. This shift presents challenges such as learning from limited labels, diverse data types, partial observations, and non-stationary conditions, all within strict latency constraints. Joint-embedding predictive architecture (JEPA) offers a potential solution by predicting missing or future representations in a latent space, rather than reconstructing raw data or relying on contrastive samples. The research provides a wireless-centric guide to JEPA for 6G intelligence, outlining how various wireless and network data can be tokenized and masked for training. A beam-management case study demonstrates that a wireless-specific target, like an auxiliary future beam-energy target during pretraining, can enhance label efficiency and robustness across different deployment scenarios compared to traditional supervised methods. The paper concludes by identifying key open challenges, including multi-timescale prediction, action-conditioned modeling, and distributed training.