【摘要】 无监督表示学习能够避免耗时且昂贵的数据标记，因此获得了极大的关注。其中的对比学习方法只需在抽象语义级别的特征空间上学习数据的区分便能使模型学习到良好的表征，被预测为人工智能的未来。无监督对比学习利用自身数据生成的伪标签作为监督信息学习样本数据的特征表达，并用于下游任务，其次核心思想是把正样本对距离拉近，正样本与负样本距离拉远。本文首先总结常用的预训练任务，然后梳理无监督对比学习在视频领域的通用框架，并对框架的各个组件及其分类进行说明，同时列举常用于评估视频表征质量的下游任务，并对其性能进行分析。之后对视频理解中的对比学习进行举例说明，最后讨论对比学习所面临的挑战和未来研究方向。更多还原

【Abstract】 Unsupervised representation learning can avoid time-consuming and expensive data tagging, so it has attracted great attention. The contrastive learning method only needs to learn data discriminationin the feature space at the abstract semantic level to learn good representation, and is predicted to be the future of artificial intelligence.Unsupervised contrastive learning uses the pseudo-labels generated by its own data as supervised information to learn the feature expression of sample data and apply them to downstream tasks. Its core idea is to shorten the distance between positive samples and expand the distance between positive samples and negative samples.In this paper, the commonly used pre-training mission is summarized firstly. Then the unsupervised contrastive learning common framework in the field of video is combed, and the various components of the framework and the classification of components are expounded. At the same time the downstream tasks commonly used to evaluate the quality of video representation are listed and their performance is analyzed.Examples are given to illustrate contrastive learning in video comprehension. Finally, challenges and future research directions of contrastive learning are discussed.更多还原