LatentWave: JEPA Pretraining for Wireless Foundation Models
Researchers have introduced LatentWave, a new wireless foundation model that utilizes a Joint-Embedding Predictive Architecture (JEPA) for pretraining. Unlike previous methods that focused on reconstructing low-level signal details, LatentWave learns more transferable representations by predicting masked regions in a latent space. This approach, which incorporates per-channel patch embeddings and stochastic channel sampling, allows the model to handle variable antenna counts and diverse wireless configurations effectively. Evaluations on tasks such as RF signal classification and 5G positioning demonstrate its improved out-of-the-box performance compared to masked-modeling baselines. AI
IMPACT Introduces a novel pretraining approach for wireless foundation models, potentially improving transferability and usability across diverse wireless tasks.