Neural Networks Against (and For) Self-Training: Classification with Small Labeled and Large Unlabeled Sets