面向深度网络的小样本学习综述

深度学习以数据为驱动，被广泛应用于各个领域，但由于数据隐私、标记昂贵等导致样本少、数据不完备性等问题，同时小样本难于准确地表示数据分布，使得分类模型误差较大，且泛化能力差。为此，小样本学习被提出，旨在利用较少目标数据训练模型快速学习的能力。系统梳理了近几年来小样本学习领域的相关工作，主要整理和总结了基于数据增强、基于元学习和基于转导图小样本学习方法的研究进展。首先，从基于监督增强和基于无监督增强阐述数据增强的主要特点。其次，从基于度量学习和基于参数优化两方面对基于元学习的方法进行分析。接着，详细总结转导图小样本学习方法，介绍常用的小样本数据集，并通过实验阐述分析具有代表性的小样本学习模型。最后总结现有方法的局限性，并对小样本学习的未来研究方向进行展望。

Survey on few-shot learning for deep network

Deep learning is widely used in various fields, which is data-driven. Data privacy and expensive data labeling or other questions cause samples absent and data incompleteness. Moreover, small samples cannot accurately represent data distribution, reducing classification performance and generalization ability. Therefore, few-shot learning is defined to achieve fast learning by utilizing a small target samples. This paper systematically summarized the current approaches of few-shot learning, introducing models from the three categories: data augmentation-based, meta-learning based, and transduction graph-based. First, it illustrated the data augmentation-based approaches according to supervised and unsupervised augmentations. Then, it analyzed the meta-learning based approaches from metric learning and parameter optimization. Next, it elaborated transduction graph-based approaches. Eventually, it introduced the commonly few-shot datasets, and analyzed representative few-shot learning models through experiments. In addition, this paper summarized the challenges and the future technological development of few-shot learning.