We prove divergence results for gradient flows for deep neural networks with analytic activation and polynomial target functions.