Explainable Neural Networks with Guarantees: A Sparse Estimation Approach