Mutual Exclusivity Loss for Semi-Supervised Deep Learning