Improved prototypical network for active few-shot learning

Different from deep learning with large scale supervision, few-shot learning aims to learn the samples’ characteristics from few labeled examples. Apparently, few-shot learning is more in line with the visual cognitive mechanism of the human brain. In recent years, few-shot learning has attracted more researchers’ attention. They all assume that each category only contains a few labeled samples and without unlabeled samples during the few-shot training. However, in reality, each visual categories also includes some unlabeled samples which store rich semantic information. Labeling all unlabeled samples is time-consuming and laborious. Therefore, we combine the few-shot learning and active learning together, and propose an active few-shot learning model (AC-FSL) based on prototype network. This model includes two core modules: few-shot classification module and loss prediction module. The former is used for classification task under few labeled samples, while the latter is applied to choose some unlabeled samples of high value for labeling, and then assist for the few-shot classification task. We conduct extensive experiments on two image benchmark datasets. The results show that the proposed model AC-FSL can effectively improve the classification performance of current methods on few-shot setting.