基于概率图模型的大规模视觉皮层脉冲神经网络计算机理研究

Large-scale Spiking Neural Network (SNN) is the basis of cognition and perception. The study of large-scale Spiking Neural Networks can not only reveal the mechanism of cognition and perception, but also provide new means for the development of neuroscience and computer vision. This project focuses on the weakness of the existing models, that is, unknown representation mechanism, small-scale model and lack of learning methods. Besides, we provide a new idea to model large-scale Spiking Neural Network based on the probabilistic graphical model. The contents of this project may involve the following issues, probabilistic neural codes, the computation of neural population based on probabilistic sampling, Large-scale SNN based on hierarchical Bayesian model, a computational model for visual processing based on SNN. Firstly, breaking the limitation of conventional neural codes, an effective neural code in line with the dynamic characteristics of neurons is provided for better representation. Secondly, by exploring the emergency and transmission of neural signals, fast distributed inference and learning algorithms are designed for effective information processing. Thirdly, to study the computation mechanism in the visual cortex, a new probabilistic framework for large-scale SNN is built. Finally, this probabilistic framework will be applied in computer vision to achieve highly accurate object recognition. The achievements of this project will improve the development of neuroscience and computer vision.

大规模脉冲神经网络（SNN, Spiking Neural Network）是人脑认知感知的基础。研究大规模脉冲神经网络计算机理不仅可以揭示人脑的认知感知机理，而且为神经科学、计算机视觉等学科的发展提供新的手段。针对现有模型存在的表达机理不明确、网络规模较小、学习算法不匹配的问题，本课题拟提出一套概率图模型来实现SNN网络的建模。研究内容包括：研究神经元信息的概率编码，基于随机采样的神经元群计算机理，基于分层贝叶斯网的大规模SNN连接机制，基于大规模SNN的视觉皮层计算模型。突破传统神经元编码方法，提出符合神经元复杂动力特性的高效信息编码；探索脉冲序列在大脑皮层发放与传导机制，设计分布式快速推理与学习算法；挖掘大脑视觉皮层的信息处理机制，构建大规模SNN的概率计算框架。最后将此计算框架应用于计算机视觉问题，实现高精度的物体识别。研究成果将推动神经科学与计算机视觉领域的共同进步。

本项目主要研究脉冲神经网络的计算机理以及大规模网络的构建。脉冲神经网络具有随机性、网络变结构性、时空信息耦合性以及大规模性等特点，这些特点即是大脑复杂性的表现也是智能的基础。如何利用这些特点构建高效的脉冲神经网络计算机理是目前面临的主要困难。目前电生理实验证明大脑皮层处理信息的机理之一是贝叶斯原理。本项目借鉴采样贝叶斯推理、变分贝叶斯推理以及贝叶斯参数估计等技术，从局部电路以及大规模全局网络两个尺度研究脉冲神经网络的计算机理。在此基础上，本项目提出了脑启发的大规模脉冲神经网络模型，构建了适合于脑启发神经网络的软件仿真平台, 实现了高效的决策学习、持续学习以及多任务学习等智能任务。本项目目前发表或录用SCI期刊论文13篇，会议论文9篇，其中包括两篇顶级国际会议：NeurIPS论文1篇和ICML论文1篇。申请7项国家发明专利。