零样本学习研究进展

近几年来,深度学习在机器学习研究领域中取得了巨大的突破,深度学习能够很好地实现复杂问题的学习,然而,深度学习最大的弊端之一,就是需要大量人工标注的训练数据,而这需要耗费大量的人力成本.因此,为了缓解深度学习存在的这一问题,Palatucci等于2009年提出了零样本学习(Zero-shot learning).零样本学习是迁移学习的一种特殊场景,在零样本学习过程中,训练类集和测试类集之间没有交集,需要通过训练类与测试类之间的知识迁移来完成学习,使在训练类上训练得到的模型能够成功识别测试类输入样例的类标签.零样本学习的意义不仅在于可以对难以标注的样例进行识别,更在于这一方法模拟了人类对于从未见过的对象的认知过程,零样本学习方法的研究,也会在一定程度上促进认知科学的研究.鉴于零样本学习的应用价值、理论意义和未来的发展潜力,文中系统综述了零样本学习的研究进展,首先概述了零样本学习的定义,介绍了4种典型的零样本学习模型,并对零样本学习存在的关键问题及解决方法进行了介绍,对零样本学习的多种模型进行了分类和阐述,并在最后指明了零样本学习进一步研究中需要解决的问题以及未来可能的发展方向.

Research and Development on Zero-Shot Learning

In recent years, deep learning has made great breakthroughs in the field of the machine learning. The application of deep learning can be especially useful for coping with some complicated problems. However, one of the biggest drawbacks of the deep learning is that it requires a great amount of manual data annotation, this issue requires a lot of labor costs. Therefore, in order to alleviate the burden of deep learning, Palatucci et al. proposed zero-shot learning in 2009. Zero-shot learning is a special scenario of transfer learning. In the zero-shot learning progress, the samples of training and test classes do not intersect. It is necessary to complete the training by realizing the knowledge transfer between training class and test class, in order to successfully identify the label of the test class instance. The significance of zero-shot learning lies not only in the identification of difficult-to-annotate, but also in the fact that this method simulates the human cognitive process of objects that have never been seen before, so the study of zero-shot learning will contribute to the advancement of human cognitive science research. In view of the application value, theoretical significance and future development potential of zero-shot learning, this paper systematically reviews the research progress of zero-shot learning. First, the definition of zero-shot learning is summarized, then we introduce four typical zero-shot learning models, and the systematic problems in the zero-shot learning and the solution methods are introduced. We have categorized and elaborated on the multiple models of zero-shot learning. At the end, we pointed out the problems that need to be solved in the further study of zero-shot learning and possible future development directions.