¿µ¹®³»¿ë (English Abstract) |
Semi-supervised learning(SSL) produces an efficient way to exploit unlabeled data. There are a number of semi-supervised learning algorithms that can be categorized into algorithm families orthogonal to each other. Our algorithm, Hard MixMatch, is a combination of commonly existing SSL algorithms. Hard MixMatch focuses on transforming an unlabeled data¡¯s target distribution to a pseudo label. Hard MixMatch performs not only sharpening the distribution but pseudo labeling based on a complicated architecture, having RandAug, Mixup, and Real MixMatch components. Hard MixMatch has also shown an improvement in accuracy in comparison to recent SSL, MixMatch.
|