One Shot Learning with Margin

Xianchao Zhang, Jinlong Nie, Linlin Zong, Hong Yu, Wenxin Liang

Abstract: One shot learning is a task of learning from a few examples, which poses a great challenge for current machine learning algorithms. One of the most effective approaches for one shot learning is metric learning. But metric-based approaches suffer from data shortage problem in one shot scenario. To alleviate this problem, we propose one shot learning with margin. The margin is beneficial to learn a more discriminative metric space. We integrate the margin into two representative one shot learning models, prototypical networks and matching networks, to enhance their generalization ability. Experimental results on benchmark datasets show that margin effectively boosts the performance of one shot learning models.