脑宏观功能网络中的高阶相互作用及自组织临界性对记忆与识别性能的影响

	脑宏观功能网络中的高阶相互作用及自组织临界性对记忆与识别性能的影响
	黄旭辉1,2
	2016-08
关键词	静息态功能磁共振成像功能连接两两相关性短时程突触可塑性蓄水池计算记忆模式识别
页数	87
中文摘要	脑科学与类脑智能系统是当前国际非常重要的科技前沿，其能有利于提升人类健康水平和发展新一代人工智能技术。在本文中，我们主要研究了两个课题：关于宏观脑网络相互作用的基本原理和自组织临界性对类脑智能系统的功能益处。这两个课题的出发点及主要结论简述如下。（一）系统神经科学的一个重大挑战是阐释不同尺度下巨量神经元之间的相互作用是如何组织，以调节知觉、运动与认知等。在过去的二十年里，研究不同脑区之间的相互作用通常基于功能连接（Functional Connectivity，简称FC）分析，FC分析是完全基于测量脑区活动的两两之间相关性。然而，理解许多复杂系统的行为需要考虑超越两两相互作用的高阶相互作用（Higher-Order Interactions，简称HOIs），它对解释系统行为至关重要。迄今为止，宏观脑网络是否存在HOIs以及HOIs如何影响大脑的活动在很大程度上都没有被认识清楚。为了解决这些疑问，我们分析了人类被试在静息态下fMRI技术记录的六个典型功能网络的血氧信号。我们发现在不同情况下——不同网络尺寸、不同拓扑结构、不同的两两相关性程度以及是否有全局信号——真正HOIs在网络内与网络间都很弱。为了理解弱HOIs背后的潜在机制，我们建立了一个受实验数据约束的网络动力学模型来生成神经信号与相应的血氧信号，通过对不同连接强度、不同网络拓扑以及不同兴奋性与抑制性连接数比的模拟数据分析。我们也发现HOIs在各种情况下都很弱。综上所述，我们的研究结果表明，弱HOIs是大脑宏观功能网络的固有属性，这意味着两两相互作用在修饰大脑活动起主导地位，同时也为广泛使用FC分析方法的有效性提供了理论支撑。此外，它也可以对设计类脑智能系统具有启发作用，即不需要考虑HOIs。（二）大脑功能的实现依赖于其保持在一个优化的状态。但目前对这一状态的特征及其与脑功能的关系还远未明了。已有大量实验提出了大脑工作在临界状态附近的猜想，实验也揭示处于临界态的神经系统具有功能优势，如实现最大的刺激响应范围、最优的信息传播以及最大的信息存储容量。另一方面，新近研究表明，通过短时程突触可塑性，神经网络能够自组织地趋向临界态并维持在其附近，即自组织临界性（Self-Organized Criticality, 简称SOC）。受此思想的启发，我们研究SOC是否可以提高递归神经网络（Recurrent Neural Networks，简称RNNs)的信息处理性能。为此，我们基于蓄水池计算（Reservoir Computing，简称RC）模型中的RNNs，将短时程突触抑制引入到不同的RNNs中，来测试它对序列记忆与模式识别任务的影响。我们发现这样一个简单的可塑性机制大大扩展了RC模型的工作的参数范围，同时它还使系统对输入噪声具有鲁棒性。这些结果源于蓄水池网络被自组织地维持在临界态附近。我们的研究结果意义在于如下两方面：一方面实证了SOC框架可以帮助优化RNNs的设计；另一方面揭示了短时程突触可塑性在神经网络功能实现的益处。这两个课题的研究成果既可以加深对大脑工作的基本原理与功能如何实现的理解，也对设计新型的类脑智能系统具有指导意义。; Brain science and brain-inspired intelligence systems currently are important frontiers in the research of science and technology, and greatly benefit for the human health and a new generation of artificial intelligence technology. In this thesis, we mainly discussed both the basic interaction principle in organizing the macroscopic brain network and the functional benefit of self-organized criticality for intelligence systems. The main ideas and results of the two projects are listed as following, respectively. (I) To explain how myriads of neuronal interactions at various scales are organized to mediate perception, motion and cognition is a major challenge for systems neuroscience. During the past two decades, interactions among different brain regions are usually examined through functional connectivity (FC) analysis, which is exclusively based on measuring pairwise correlations in activities. However, interactions beyond the pairwise level, i.e., higher-order interactions (HOIs), are vital in understanding the behavior of many complex systems. So far whether HOIs exist among brain regions and how they can affect brain’s activities remain largely elusive. To address these issues, here we analyzed blood oxygenation level-dependent (BOLD) signals recorded from six typical brain networks in human subjects during rest. We found that true HOIs both within and across individual networks were very weak, regardless of the network size, typology, pairwise correlation strength, and whether the global signal was regressed or not. To investigate the potential mechanisms underlying the weak HOIs, we analyzed the dynamics of a network model constrained by the empirical data, and also found that HOIs were generally weak within a wide range of connection strength, network topology and composition of excitatory and inhibitory connections. Taken together, our results suggest that weak HOI is an intrinsic property of brain’s macroscopic functional networks, which implies the dominance of pairwise interactions in shaping brain activities and warrants the validity of widely used FC approach. Moreover, it could enlighten the design of the brain-inspired intelligence systems without considering the HOIs. (II) The realization of various functions of the brain relies on its being organized at an optimized state. A number of previous studies have hypothesized that the brain works at or close to the critical state. It has also been reported that nervous systems that are critical have various functional advantages, such as maximized dynamic range, optimized information transmission and maximized information storage capacity. On the other hand, it has been suggested that, through short-term synaptic plasticity, neuronal dynamics in the brain are self-organized close to a critical state. Inspired by such an idea, here we investigate if and how the self-organized criticality (SOC) can improve the performance of a recurrent neural network (RNN) in information processing. To this end, we implemented short-term synaptic depression into RNNs in different models of reservoir computing (RC) and tested its effects on sequence memory as well as pattern recognition tasks. We found that such a simple form of plasticity greatly expanded the parameters range within which the RC models can perform well. In addition, it makes the system more robust to input noises. These results are achieved through dynamically maintaining the reservoir close to a critical state. Our study suggests that the SOC framework can be instrumental in optimizing the design of RNNs. In addition, these results shed new light on the functional benefits of short-term synaptic plasticity for neural networks. Results of these two projects could deepen the understanding of the basic principle of brain’ organization and the realization of various real functions, and guide the design of new brain-inspired intelligence systems.
所属项目编号	11505283 ; 2015T80154
所属项目名称	自然科学青年基金：临界态对生物神经网络学习、记忆以及模式识别能力的影响 ; 博士后第八批特别资助：临界态对生物神经网络记忆与分类能力的影响
文献类型	研究报告
条目标识符	http://ir.ia.ac.cn/handle/173211/12032
专题	脑图谱与类脑智能实验室_脑网络组研究
作者单位	1.中国科学院自动化研究所模式识别国家重点实验室 2.中国科学院自动化研究所脑网络组研究中心
推荐引用方式 GB/T 7714	黄旭辉. 脑宏观功能网络中的高阶相互作用及自组织临界性对记忆与识别性能的影响. 2016.

条目包含的文件		下载所有文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
博士后出站报告_黄旭辉.pdf（5305KB）	研究报告		开放获取	CC BY-NC-SA	浏览下载