Unlearning via Sparse Representations