视觉信息层级表征机制及其计算模型

Human beings have highly intelligent visual systems, which are much stronger than advanced artificial intelligence (AI) systems in many aspects. Human intelligence is based on a set of remarkably powerful representations. However, the details of the representing mechanism of human visual system remains unclear. Our previous study focused on dynamic objects to investigate the representing mechanism of human visual system, finding that the motion relationship of objects were represented in hierarchical trees. Accordingly, computational models were built for the hierarchical representations. Given that hierarchical tree is one kind of the prominent representations that could support human vision system to achieve rapid learning and knowledge transfer (to novel situations), this project aims to further explore the mechanism of hierarchical representations, by using psychological experiments and modeling techniques. On the basis of our previous study that revealed the mechanism of hierarchical representation for dynamic objects, this project plans to generalize this finding to static objects, in order to systematically examine the properties of hierarchical representation for variable kinds of visual information. Based on these results, we plan to further reveal the grammar rules of the hierarchical tree used by visual system, and the operations implemented on the hierarchical representations. The computational model of the hierarchical representation will be developed according to the empirical findings, so as to explain the representing mechanism of visual system by the language of mathematics, therefore proving precise descriptions for psychological theory and promoting the connection between psychology and AI. Results of this project are expected to deepen the understanding of visual intelligence by revealing and modeling how visual information is processed based on their hierarchical representations, and inspire revolutionary transform of the knowledge representation for AI systems.

人类具有高度智能化的视觉系统，即使当今最前沿的人工智能技术在许多方面仍无法企及。这种智能化主要源于人类优良的表征系统。现有研究尚未揭示其具体表征机制。申请者团队的研究以运动客体为例，发现运动关系表征存在层级结构，并构建了相应的层级模型。层级表征的良好特性，为人类视觉快速学习和灵活迁移提供了基础。据此本项目拟采用认知建模技术与心理学实验相结合的方法，考察视觉信息的层级表征机制及其加工规律。基于前期工作，申请人拟针对静态客体，进一步系统考察视觉信息的层级表征机制，构建层级树结构相对应的语法体系，并探索基于层级表征的认知加工过程。同时，本项目将构建人类视觉层级表征的计算模型，采用数学语言精确阐释视觉系统的表征规律，从而推进心理学相关理论的发展及其与人工智能的对接。本项目研究成果可望加深对人类视觉智能的理解，为视觉层级表征的心理加工过程提供计算模拟，并为人工智能构建革命性表征体系提供心理学支持。

人类认知系统拥有高效的表征体系，这是人类强大智能的基础。然而，当前研究对于心理表征具体机制的了解尚浅。因此，本项目采用认知建模技术与心理学实验相结合的方法，考察视觉信息的层级表征机制及其加工规律。研究探明了视觉客体层级表征的一般规律、结构特性和语法规则，揭示了其中从低层视觉特征抽取高层语义信息的机制，并构建了相应的层级表征计算模型。在实现预期目标的同时，还补充了客体关系工作记忆等研究，并在智能交互意图识别方面展开成果应用工作。研究成果在Journal of Experimental Psychology: General等期刊和Annual Meeting of the Cognitive Science Society 等会议上发表研究报告13篇，博士论文4篇、硕士论文8篇。这些成果有助于深入理解人类认知系统信息表征方式和操作规律，可为开发具有通用知识表征的人工智能系统提供启示。