Scheduled sampling for one-shot learning via matching network

Abstract Considering human can learn new object successfully from just one sample, one-shot learning, where each visual class just has one labeled sample for training, has attracted more and more attention. In the past years, most researchers achieve one-shot learning by training a matching network to map a small labeled support set and an unlabeled image to its label. The support set is combined by one image with the same label as unlabeled image and few images with other labels generated by random sampling. This random sampling strategy easily generates massive over-easy support sets in which most labels are less relevant to the label of unlabeled image. It leads to the limitation of matching network for one-shot prediction over indistinguishable label sets. For this issue, we propose a novel metric to evaluate the learning difficulty of support set, where this metric jointly considers the semantic diversity and similarity of visual labels. Based on the metric, we introduce a scheduled sampling strategy to train the matching network from easy to difficult. Extensive experimental results on three datasets, including mini-Imagenet, Birds and Flowers, indicate that our method could achieve significant improvements over other previous methods.