基于深度Spiking神经网络的多模态类脑模型研究

深度学习成功借鉴大脑处理信息的层级结构，在智能应用领域取得了重要突破。但是现有深度模型距离人脑的智能水平还有很大差距，仍然需要更深层次地借鉴和融合大脑在微观、介观、宏观多尺度的结构和信息处理机制。本课题综合考虑生物神经系统的脉冲时间编码方式，深度层次化信息处理结构和多脑区协同认知等类脑特性，研究具有多模态信息整合能力，支持复杂视听感知-学习-决策的深度Spiking神经网络类脑模型。具体而言，从神经编码基础理论出发，研究基于脉冲时间机制的多模态信息编码；其次，借鉴不同脑区模块间的协同处理模式，构建可以关联和处理多模态信息的深度Spiking神经网络模型；继而探索受脑启发的突触可塑性学习规则，为类脑计算模型提供高效学习的算法；最后，搭建计算平台，用实际应用验证所提模型的性能。本项研究将促进类脑计算的理论发展和提升类脑模型的实际应用能力。

Deep learning draws on the hierarchical structure of the brain's information processing mechanism and has achieved important breakthroughs in the field of intelligent applications. However, it is not enough to draw on the macroscopic neural structure of the brain to achieve a level of intelligence similar to the human brain. To achieve human-like intelligence, computational models need to be able to assimilate the brain's micro-and macro-scale structures and information processing mechanisms. This project comprehensively surveys the brain-like characteristics of the spike-timing-based encoding, deep hierarchical information processing structure, and multi-brain area collaborative cognition mechanism, and studies a multi-modal perception model based on deep Spiking neural networks. Firstly, our research anchors on the basic theory of neuromorphic coding and spike calculation and investigate the spike-timing-based encoding of multi-modal information. Secondly, we construct a multi-modal deep spiking neural network model drawing inspiration from the collaborative information processing similar to how different brain regions communicate. Thirdly, we explore brain-inspired synaptic plasticity learning rules to provide efficient learning algorithms for brain-like computing models. Finally, we build a computing platform to verify the performance of the proposed model with practical applications. This research will promote the theoretical development of brain-inspired computing and enhance the practical application of brain-inspired models.