On exponential convergence of SGD in non-convex over-parametrized learning