Analyzing Convergence Aspects of Federated Learning: More Devices or More Network Layers?

Federated learning has attracted considerable research interest to better shape the next-generation communication systems. Along this line, in this paper, different combinations of edge devices and convolutional layers of neural network are tested for global model convergence. We investigate the number of communication rounds (CRs) required to make a global model converge, when the number of convolutional layer channels and edge devices taking part in global model convergence varies. We observe the effects of additive white Gaussian noise (AWGN) on gradient vectors (GVs) that are shared with the parameter server (PS) through a wireless channel. Further, we add channel impairments and observe the CRs required to make the model converge. With higher values of noise power and channel impairments, even after exhausting the maximum number of CRs, the global model do not converges for lower number of edge devices and convolutional layer channels. However, if either or both the number of edge devices and convolutional layer channels are increased, the global model converges with substantially higher accuracy even for stronger noise and channel effects.