Text this: Closed-form interpretation of neural network classifiers with symbolic gradients