ML Interview Q Series: In the context of optimization tasks that may be convex or non-convex, does the gradient in stochastic gradient descent always direct us to the universal optimum value?