Convergence of stochastic gradient descent under a local Lajasiewicz condition for deep neural networks