The High Separation Probability Assumption for Semi-Supervised Learning