SGD introduces noise into the gradient estimate. In non-convex optimization, how can this noise be beneficial?