wireless datasets

This talk presents the Large Wireless Model (LWM), the world’s first foundation model for wireless channels. Inspired by the success of foundation models in NLP, speech, and vision, LWM is a transformer-based model pre-trained in a self-supervised fashion on large-scale diverse wireless datasets. It learns rich, universal contextualized channel embeddings (features) that potentially enhance performance across a wide range of downstream tasks. I will present the model’s architecture, its self-supervised pre-training approach, and training datasets. I will also demonstrate its gains in tasks like sub-6GHz to mmWave beam prediction, LoS/NLoS classification, and localization. These gains highlight the LWM’s ability to learn from large-scale wireless data and enable complex machine learning tasks with limited data in wireless communication and sensing systems.

Finally, we introduce an ITU AI/ML 5G competition which provides a modular setup, where participants can innovate on scenario design, feature extraction, and lightweight downstream models, pushing the frontiers of robustness, generalizability, and interpretability. By contributing improved scores and model refinements, the challenge also opens doors for discussion on formats, reproducible simulations, and alignment with 6G use cases. The outcomes are expected to influence real-world deployments, research reproducibility, and standard frameworks for wireless AI.

Learning Objectives:

1. Describe the architecture and self-supervised training approach of the Large Wireless Model (LWM).
2. Explain how LWM generates contextualized channel embeddings and how they contribute to wireless communication and sensing tasks.
3. Analyze the performance of LWM in downstream tasks such as beam prediction, LoS/NLoS classification, and localization.
4. Evaluate the role of large-scale data and foundation models in improving generalizability and efficiency in wireless AI applications.
5. Design innovative approaches for feature extraction or scenario modeling and apply them in the ITU AI/ML 5G challenge.