On the existence of global minima and convergence analyses for gradient descent methods in the training of deep neural networks