AI and the World's Hard Problems · Theoretical computer science · arXiv · 2026-10-22
Learning to Interfere in Non-Orthogonal Multiple-Access Joint Source-Channel Coding
When multiple transmitters communicate source signals such as images over a multiple access channel, conventional systems limit capacity by orthogonally allocating time and bandwidth to minimize interference. The proposed ML-aided wireless image transmission method merges compression and channel coding with a multi-view autoencoder, letting transmitters use all available channel resources simultaneously in a non-orthogonal multiple access (NOMA) scheme where the receiver recovers all images from the superposed signal and associates each with its transmitter. A progressive fine-tuning algorithm doubles the number of users each iteration, supporting up to 64 users with only a 2.4% increase in trainable parameters over a single-user model while outperforming existing NOMA-based methods across a wide range of datasets, metrics, and channel conditions.