Repeated Random Sampling for Minimizing the Time-to-Accuracy of Learning