人脑记忆机制启发下的视觉信息表达、存储和提取方法研究

本项目将借鉴认知神经科学的最新研究成果，结合模式识别和机器学习理论对视觉信息在人脑中的表达、存储和提取方法进行研究，建立人脑视觉信息的情景和语义记忆模型，并将其用于大规模视觉数据的压缩、存储、分类和检索中。具体地，首先基于卷积神经网络对大脑视皮层腹侧视觉通路各区域神经元响应特性进行模拟，提出一种新的层次式视觉感知信息的不变性稀疏表达模型；然后建立一种新的标签嵌入字典学习算法对大脑内侧颞叶中鼻周皮层进行建模，用于模拟人脑语义记忆过程，并通过构建模式关联与竞争神经网络模型对海马体内的齿状回进行建模，以实现模式分离和情景记忆的快速存储；而在记忆提取时首先通过构建模式关联和自联想神经网络对海马CA3建模以对不同提取线索进行关联和模式补全，然后对海马CA1建模将解码后的视觉模式经内嗅皮层反向映射到鼻周皮层进行语义提取。

In this project, based on the theories of pattern recognition and machine learning and the insights gleaned from the latest research results of cognitive neuroscience, we will investigate and study the modeling of representation, storage, and retrieval of visual information in human brain and establish corresponding episodic as well as semantic memory models and apply them to the compression, storage, classification, and retrieval of big visual data. Firstly, the convolutionary neural network (CNN) is applied for simulating the response of the neurons of each brain area in the ventral stream of the visual cortex and establishing a new hierarchical layered sparse invariant representation model for visual information. Then, a new label embedded dictionary learning algorithm is established for modeling the perirhinal cortex in medial temporal lobe, which can be applied to simulate the semantic memory in brain. Additionally, pattern association and competitive neural network models are constructed for modeling the dentate gyrus in hippocampus, which can be used to implement pattern separation and fast episodic memory. While regarding the memory retrieval, the pattern association and autoassociation neural networks are adopted first for modeling the pattern completion of different incomplete retrieval cues in CA3, and then the CA1 is modeled for backprojecting the decoded visual pattern into the perirhinal cortex via the entorhinal cortex for semantic memory retrieval.