Predicting the outputs of finite deep neural networks trained with noisy gradients