Learning Better Representations with Fewer Labels