现代信息社会多源视频大数据无处不在，且具有海量庞杂、多源异质特点。传统机器学习方法对其挖掘发现互联关系、提取有效信息和识别能力不足。参考大脑神经元多层深度互联特性，现有深度学习方法在视频图像处理和识别方面取得了长足进展，但仍然存在网络连接方式单一、缺乏联想记忆能力、对多源视频信息处理能力弱等问题。随着神经科学研究的进一步发展，受脑神经元双向连接、联想记忆匹配选择处理输入视频信息、对多源视频信息分治且融合等机制的启发，本项目研究面向视觉认知的工神经元多样性连接的网络模型，探索具有记忆联想和视觉信息选择特性的深度学习方法，以及多源视频数据分布式处理与融合机制，创建脑启发式深度学习网络，以期突破现有深度学习网络对多任务自适应、对先验信息利用、以及对多源数据处理能力不足的瓶颈，有效提高深度学习对多源视频大数据的处理能力，并丰富机器学习理论与方法。

Multi-source video big data (MSVBA) is ubiquitous in modern information age. The vast amount of complex and heterogeneous MSVBA prevented the application of traditional machine learning methods from extracting useful information for scene recognition and understanding. Inspired by the preliminary discovery on brain science, the deep learning (DL) technology has achieved a great success in image/video processing and recognition. However, it is still in its infancy, in that the working DL models suffer from several problems, i.e., simple connection and memory missing, which prevent multi-source video information from being processed efficiently. Recent neuroscience discovery on brain circuits, i.e. multi-direction connections among neurons, especially that subserving visual recognition, i.e. video information selection under the guide of associate memory, has shed light on developing more practical and DL theories and models. In the current project, we attempt to design a novel diversity connection based DL network, which combined the aforementioned features of the brain. Moreover, we attempt to further improve the DL network which can effectively process multi-source video information. The proposed DL network is adaptive to multi-tasking. It can also utilize prior information and effectively process multi-source information. Hopefully, the proposed work will improve the processing capacity of machine learning on multi-source big data, which may shed new light on machine learning theories and applications.