Learning-Based Low-Rank Approximations