Improving Few-Shot Learning Through Multi-task Representation Learning Theory