ZECCHIN Matteo - Communication systems
Date: - Location: Eurecom
For AI to be successfully deployed on the RAN, an important requirement is the ability to quickly adapt to changing environmental conditions based on limited contextual information. As a notable example, an AI-based wireless receiver should be able to update its internal operations based on limited pilots, ensuring satisfactory performance despite time-varying channel conditions. In this talk, we explore and understand the potential of in-context learning, a form of meta-learning that requires no explicit model updates to adapt the operation of AI modules. Specifically, we propose the use of pre-trained sequence models for implementing equalization over non-linear MIMO channels in a cell-free system via in-context learning and showcase the superiority of ICL compared to localized and centralized MMSE equalizers in cases of limited front-haul capacity.
