Approximate Gaussianity Beyond Initialisation in Neural Networks