预测计算神经环路的数学建模研究

Our brain is the major organ that receives and processes various incoming sensory stimuli in the environment. Understanding the general principle of information processing in our brain is a fundamental problem in neuroscience. Traditionally, this information processing is considered to a bottom-up process that represents sensory stimuli in a hierarchical manner.In contrast, a new framework, called predictive computation, emphasizes the top-down rather than bottom-up nature of information processing in our brain. Despite there is growing evidence for this new framework, its underlying neural circuit mechanism is yet to be elucidated. In this project, by incorporating the latest physiology, connectivity and synaptic plasticity data, I will perform mathematical modelling of spatiotemporal dynamics of neural circuit that consists of various types of neurons in mouse cerebral cortex. In close collaboration with leading experimental labs, the generated prediction from this model will be tested in the experiments with the state-of-the-art techniques. Through this iterative interaction between mathematical modelling and experiments, this project aims to build the first mathematical model for predictive computation in mouse cerebral cortex, thereby elucidating the neural mechanism underlying predictive computation and potentially providing insights to brain-inspired machine learning algorithms.

大脑是人类接收处理外界各种复杂信息的主要机体器官。解析大脑信息处理过程的一般原理是神经科学中的基本问题。传统上,大脑对外界信息的处理被认为是一个自下而上的由简单表征过渡到高级表征的过程。而预测计算观点则认为大脑在自上而下地生成对外界信息的预测的同时,自下而上地处理外界信息与该预测之间的差别。这种预测计算观点在近些年得到了一些实验证据的支持,但对其神经环路机理的研究才刚刚起步。本项目拟从最新的神经形态、神经电生理、神经元连接组及突触可塑性等实验数据出发,对小鼠大脑皮层中由多种不同类别神经元组成的神经环路的时空动力学进行数学建模,进而产生的重要预测将在相关合作实验室进行检验并由此往复。通过这种数学建模与实验紧密交互的方式,本项目将有机会建立国际上第一个预测计算的大脑皮层神经环路动力学数学模型,揭示预测计算的神经机理,从而提示大脑信息处理过程的一般原理,启发类脑智能高效算法。

人们对大脑如何存储序列信息仍不甚清楚。通过综合使用猕猴序列学习行为范式、双光子钙成像技术及高维神经活动降维技术，我们发现不同次序的信息在记忆保持期存储于由前额叶神经元活动组成的高维状态空间的不同子空间中，从而揭示了序列工作记忆在猕猴前额叶表征的几何结构。进一步地，通过分析不同子空间和单神经元的关系，该研究工作推翻了基于单神经元发放性质的序列工作记忆模型，提示应从神经元群体编码角度重新构建序列加工计算模型。