Interpretable conservation law estimation by deriving the symmetries of dynamics from trained deep neural networks | Synapse