Don't Wait, Just Weight: Improving Unsupervised Representations by Learning Goal-Driven Instance Weights