Wireless FEC-less Frame Transmission for Model Aggregation of Distributed Learning Systems

This paper examines a reduced necessity for forward error correction (FEC) within the application context of federated learning (FL). The common usage of FEC to reduce bit errors imposes a cost of reduced communication speeds and increased complexity. Yet, FL is shown to have some error-resilience when training. Therefore, showing a system alignment of bit error rate flexibility within a FL application can drive towards lower encoder/decoder complexity and higher communication rates. By modifying a standard physical layer to remove FEC, a wireless system can better meet the needs of FL. The resulting modifications give priority to bits based on their importance to the model's accuracy without sacrificing communication rate or integrity of the training process. This approach and demonstration establishes a path towards a more balanced importance to each system layer abstraction removing critical bottlenecks of complexity and improving overall system performance.