Asymptotics of Gradient-based Neural Network Training Algorithms