Physics-based Meta Learning for Channel Transformation

Adapting machine learning (ML) solutions in the PHY layer to new wireless configurations is hindered by the prohibitive cost of data collection for each unique configuration. While existing methods attempt to transform data from a reference to a target configuration, their reliance on large, hard-to-obtain paired datasets is a significant bottleneck. We propose a physics-based transformation framework that leverages a parametric latent space in which channel transformations can be mapped to relatively simple translations. Our model pre-trains on readily available, unpaired data to learn this space, then fine-tunes with only a fraction of paired labeled data to learn the required translation. We demonstrate the effectiveness of our framework through experiments on different modalities of channel transformation and through improvements in downstream tasks relevant to wireless communication.