Global Convergence Analysis of Local SGD for Two-layer Neural Network without Overparameterization