摘要:小样本学习旨在通过少量样本学习到解决问题的模型.近年来，在大数据训练模型的趋势下，机器学习和深度学习在许多领域中取得了成功.但是在现实世界中的很多应用场景中，样本量很少或者标注样本很少，而对大量无标签样本进行标注工作将会耗费很大的人力.所以，如何用少量样本进行学习就成为目前人们需要关注的问题.系统地梳理了当前小样本学习的相关工作，具体来说介绍了基于模型微调、基于数据增强和基于迁移学习这3大类小样本学习模型与算法的研究进展；将基于数据增强的方法细分为基于无标签数据、基于数据合成和基于特征增强这3类，将基于迁移学习的方法细分为基于度量学习、基于元学习和基于图神经网络这3类；总结了目前常用的小样本数据集和代表性的小样本学习模型在这些数据集上的实验结果；随后对小样本学习的现状和挑战进行了概述；最后展望了小样本学习的未来发展方向.



Abstract:Few-shot learning is defined as learning models to solve problems from small samples. In recent years, under the trend of training model with big data, machine learning and deep learning have achieved success in many fields. However, in many application scenarios in the real world, there is not a large amount of data or labeled data for model training, and labeling a large number of unlabeled samples will cost a lot of manpower. Therefore, how to use a small number of samples for learning has become a problem that needs to be paid attention to at present. This paper systematically combs the current approaches of few-shot learning. It introduces each kind of corresponding model from the three categories: fine-tune based, data augmentation based, and transfer learning based. Then, the data augmentation based approaches are subdivided into unlabeled data based, data generation based, and feature augmentation based approaches. The transfer learning based approaches are subdivided into metric learning based, meta-learning based, and graph neural network based methods. In the following, the paper summarizes the few-shot datasets and the results in the experiments of the aforementioned models. Next, the paper summarizes the current situation and challenges in few-shot learning. Finally, the future technological development of few-shot learning is prospected.



PDF全文下载地址:

http://jos.org.cn/jos/article/pdf/6138