eduzhai > Applied Sciences > Engineering >

Building One-Shot Semi-supervised (BOSS) Learning up to Fully Supervised Performance

  • Save

... pages left unread,continue reading

Document pages: 12 pages

Abstract: Reaching the performance of fully supervised learning with unlabeled data andonly labeling one sample per class might be ideal for deep learningapplications. We demonstrate for the first time the potential for buildingone-shot semi-supervised (BOSS) learning on Cifar-10 and SVHN up to attain testaccuracies that are comparable to fully supervised learning. Our methodcombines class prototype refining, class balancing, and self-training. A goodprototype choice is essential and we propose a technique for obtaining iconicexamples. In addition, we demonstrate that class balancing methodssubstantially improve accuracy results in semi-supervised learning to levelsthat allow self-training to reach the level of fully supervised learningperformance. Rigorous empirical evaluations provide evidence that labelinglarge datasets is not necessary for training deep neural networks. We made ourcode available at this https URL to facilitate replicationand for use with future real-world applications.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×