Sign Gradient Aggregation for Wireless Federated Learning Using Bussgang LMMSE Estimation
Seunghoon Lee, Chanho Park, Namyoon Lee · 2021 International Conference on Information and Communication Technology Convergence (ICTC) · 2021
The key challenge of wireless federated learning is to reduce the communication cost for aggregating local gradients computed with personally generated data. The computational complexity for estimating local gradient parameters at an access point can be excessive since the cost increases with the model size and the number of client devices. To alleviate this, we propose a computationally efficient gradient aggregation method for wireless federated learning by employing the Bussgang linear minimum mean squared error (BLMMSE) estimator. The key idea of BLMMSE aggregation is to model the gradient parameters as independent and identically distributed Gaussian random variables. Especially, we derive the closed-form BLMMSE solution for the one-bit quantization and demonstrate that the proposed method maintains the learning performance in the level of soft-signSGD algorithm whose aggregator is the exact minimum mean squared error estimator.