自监督学习本质的有效性探究、论证及拓展增强

近年，自监督学习(SSL)作为一种新的无监督学习范式渐受关注，并已广泛应用于计算机视觉和自然语言处理等众多领域。SSL旨在减少这些现实领域中对大量标注数据的依赖。其关键是创建一个免人工标注的前置任务，借助它，SSL能有效学得有益于下游任务的特征表示。现有SSL主要聚焦于如何设计新的前置任务，而罕有对其工作有效本质特性的深入探究，这使自监督信号的设定更多依据直觉和启发性，导致部分无效预置，因而亟需提供原则性的设计指导。我们初步分析发现，如从多视图角度入手，SSL引入的自监督信号能纳入我们早期提出的单视图数据的多视图化学习框架。据此，本项目旨在已有工作基础上，尝试借鉴多视图学习方法论探究和论证SSL本质的有效性所在，以弥补现有工作的不足。为此整个工作将围绕其基础问题展开方法论验证、本质有效的理论分析和拓展增强等。所获成果可望在推进该范式和多视图学习自身等研究方面提供借鉴。

In recent years, as a new unsupervised learning paradigm, self-supervised learning (SSL) is receiving increasingly more attention. It aims to reduce the challenging demand for a large number of labeled data and has been widely used in many fields such as computer vision and natural language processing. The key to SSL is to create a pretext task without annotation of data, with its help, SSL can effectively learn the feature representation beneficial for downstream tasks. Existing SSL studies mainly focus on how to design new pretext tasks, while few on its essential effectiveness mechanism, making the setting of self-supervised signals more intuitive and heuristic, thus leading to some invalid presets, so a principled guidance of its design is urgently needed. From the multi-view perspective, we have preliminarily found that the self-supervised signals introduced by SSL can be accommodated into our previously-proposed multi-view learning framework of single-view data. Thus our project based on existing works aims to explore and verify the effectiveness essence of SSL with the multi-view learning methodology, so as to compensate the substantial deficiencies of existing studies. To the end, we will focus on its basic problems to carry out methodological verification, theoretical analysis of essential effectiveness and extension enhancement. Our final achievements can be desirable to not only promote the research of this paradigm but also provide reference for related research including multi-view learning itself.