Making Classical Machine Learning Pipelines Differentiable: A Neural Translation Approach