SIA OpenIR  > 机器人学研究室
哺乳动物空间导航的神经计算模型研究
Alternative TitleNeural Computational Models of Mammalian Spatial Navigation
曾太平1,2
Department机器人学研究室
Thesis Advisor斯白露
Keyword空间认知与导航 计算神经科学 自主移动机器人 认知地图构建 类脑智能
Pages154页
Degree Discipline模式识别与智能系统
Degree Name博士
2019-05-25
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract本文的研究内容主要是针对哺乳类动物大脑空间认知功能中所涉及的导航理论进行研究。根据小鼠大脑中内嗅皮层-海马神经回路的最新神经生物学实验结果,探索动物空间认知,记忆及导航的机制和相关细胞特性,包括联合网格细胞,联合头朝向细胞,速度细胞,网格细胞,头朝向细胞,位置细胞,邻里细胞,并根据解剖学结构以及认知功能,主要开展以下研究:I. 为了研究哺乳类动物在环境探索的过程中如何同时编码自身的运动信息和空间位置,提出了一种空间与运动联合编码的认知地图构建模型,能够对运动信息和感知信息进行融合,并在环境中形成稳定的空间编码。对内嗅皮层-海马回路中的神经动力学过程进行了建模,能够利用来自局部视图细胞的输入,所建模的神经动力学过程能够作为一个通用的机制进行误差校正,并实现神经网络编码的模式完成功能。通过实验验证,在大尺度的户外环境中,此认知地图构建模型仅仅采用视觉信息就能够构建出连贯一致的认知地图。II. 为了进一步探索多模感知融合在哺乳动物空间导航过程中的潜在神经动力学机制,提出一种多模感知融合的贝叶斯吸引子网络模型,对耳前庭信息和视觉感知信息在头朝向细胞环形吸引子网络和网格细胞圆环形吸引子网络中融合的神经动力学过程进行了建模,编码了动物在运动过程中的位置和头方向。而这种在空间认知系统中存在的贝叶斯推理机制,可能以同样的编码形式存在与其他的脑区。根据提出的贝叶斯吸引子网络模型,将其实现为高效的类脑移动机器人SLAM系统。在大尺度的环境和典型的机器人环境中对算法进行了测试,能够正确的构建出所探索环境的地图,算法在大尺度的环境具有很强的抗噪和抗不确定性的能力。III. 为了研究哺乳类动物如何利用认知地图对所探索空间进行高效的编码,在环境中长期工作并控制认知地图大小的增长,提出一个简单,实用而高效的紧凑认知地图构建方法。从实验神经科学单细胞活动记录中发现的邻里细胞,所启发引入邻里域的概念对环境进行划分,采用地形定向,利用航向信息实现紧凑的认知地图构建。此算法能够缓和的权衡取舍认知地图的精度和计算代价之间的关系。所引入的邻里域,能够用于调节所构建认知地图的稀疏度。IV. 为了从理论的角度解释网格细胞局部锚定的发放模式是如何形成全局的发放模式,提出一个从局部锚定编码到全局一致编码的转换模型,其能够将头朝向细胞和网格细胞的局部锚定的发放模式转换成全局一致的发放模式,能够对所探索环境实现全局一致空间编码。全局一致发放模式编码为大脑中存在高效,精确,统一的全局一致的空间编码提供理论支持,对于探究导航的生物性基础具有重要意义。提供了理论性依据支持基于经验校正的全局编码,描述了内嗅皮层路径积分位置估计和海马外部位置信息联想学习之间的相关作用,有可能进一步加速对空间认知与记忆的理解,更一般的来讲,对理解人类情景记忆提供假设和理论基础。V. 为了更好的理解内嗅皮层-海马神经回路在空间认知与空间导航上的作用,对从内嗅皮层到齿状回所产生稀疏编码的模式完成功能进行建模,以解决动物在环境探索过程中由于所处周围环境相似性引起的不确定性,区分所处的不同位置。提出一种从内嗅皮层第二层背侧到腹侧多个不同网格间隔网细胞模块到齿状回颗粒细胞的权重连接模型,此连接在齿状回中形成类似局部敏感哈希编码的编码方式,权重利用海扁竞争学习和控制齿状回颗粒细胞稀疏度的方式,在齿状回中形成赢家通吃的机制,从而产生稀疏发放位置编码方式,打破动物在环境感知过程中的不确定性。这种类似局部敏感哈希编码方式不仅存在于空间认知实现对环境感知信息消除不确定性,而且很有可能作为一种通用的神经机制存在于多个不同的脑区以及不同物种的大脑中。VI. 为了解释内嗅皮层中网格细胞模块失活后位置域扩展问题,提出一种基于贝叶斯推理的傅里叶空间频率假设模型。贝叶斯傅里叶空间频率假设模型成功解释了位置域扩展问题,并预测网格细胞模块活动峰对齐位置域形成的结果。更进一步的,根据内嗅皮层-海马神经回路的解剖连接和认知功能,对内嗅皮层-海马神经回路进行建模,探索动物导航过程中的空间导航和情景记忆认知功能,提出空间记忆索引理论来解释空间认知与空间导航,以及情景记忆。在动物探索新环境的过程中,新的记忆如何在新皮质,内嗅皮层-海马神经回路中建立连接。旧的记忆如何在动物再次访问时,再次被唤起,如何用于空间导航和环境认知。本文对哺乳类动物空间认知与导航的内在神经机制进行探索,在脑认知基础研究方面,用吸引子网络解释了在环境探索时如何同时编码运动与位置;贝叶斯吸引子网络实现多模感知信息融合;邻里细胞实现高效稀疏的认知地图编码;为全局一致的空间编码提供理论依据;位置细胞形成的类似局部敏感哈希编码能够打破感知过程的不确定性;用贝叶斯傅里叶空间频率假设解释位置域的扩展,并提出空间记忆索引理论解释空间认知与导航,以及情景记忆。在类脑智能研究方面,对所提出的理论假设以及计算模型在机器人平台上进行了验证,并启发了新的类脑移动机器人算法。所提出的计算模型能够促进更好的理解空间认知的神经机制,帮助建立更加精确的大脑模型,促进更好的理解大脑;同时提供更多的可能性建立更加智能,更加可靠的导航系统。
Other AbstractThe research contents mainly focus on navigation theory involving in spatial cognition of mammalian brain. According to recent neurobiological experiments in the entorhinal-hippocampal neural circuit of rat brain, it explores the neural mechanism of spatial cognition, memory and navigation, and characteristics of spatial coding related neurons, including conjunctive grid cells, conjunctive head direction cells, speed cells, grid cells, head direction cells, place cells, neighborhood cells. Based on the anatomical structure and cognitive functions, the following studies are mainly performed: I. In order to understand the neural mechanism of conjunctive representation of movement and space during mammalian exploration in the environment, a cognitive computational model is proposed to simultaneously represent movement and space information, integrate this two information, and form a stable spatial representation of explored environments. The neural dynamic in the entorhinal-hippocampal circuit is modeled, which is able to utilize the input from local view cells. The modeled neural dynamic can be considered as a general mechanism of error correction, and implement the pattern completion function of neural coding. The performance of the proposed cognitive mapping model is demonstrated in the large-scale outdoor environment. It is capable of producing a coherent semi-metric topological map only using a monocular camera. II. A Bayesian attractor network is proposed to explore the neural mechanism of multiple sensory integrations in the mammalian brain during the navigation process. Vestibular and visual cues are integrated on the ring attractor neural manifold of head direction cells and the torus attractor neural manifold of grid cells representing head direction and position of mammals, respectively. The Bayesian inference mechanism is likely to be a general principle and model for multisensory integration existing in other neural systems, besides the spatial navigation neural system. The proposed Bayesian attractor model is implemented as an efficient brain-inspired robot navigation system. It is demonstrated in the large-scale outdoor environment and the typical robot maze. Coherent cognitive maps are generated by the proposed algorithm. It is robust to noise and uncertainty existed in the natural environments during mammalian navigation process. III. To understand how efficient spatial codings are achieved by cognitive maps and control the growth of the size of the cognitive map, a pragmatic, simple, efficient compact cognitive mapping solution is proposed. According to the discovery of neighborhood cells found in single cell activity recording experiments, the concept of a neighborhood field is introduced to segment distinct parts of the explored environment with topological orientation. Course information is utilized to implement a compact cognitive mapping system. The proposed algorithm allows to gently trade off accuracy for computational cost including computational time and memory footprint. The introduced concept of neighborhood field is able to adjust the sparsity of established cognitive map. IV. To provide a theoretical explanation of how grid cells form a global coherent firing pattern, a transformation model is proposed to transit local anchored firing patterns to global coherent firing patterns, both for head direction cells and grid cells. It is capable of supporting the universal spatial metric for mammals spatial navigation in all environments. The global coherent firing pattern provides a piece of possible theoretical evidence for an efficient, accurate, universal spatial representation. It is crucial for studying the neural mechanism of mammalian navigation. Experience-dependent interactions between the hippocampus and entorhinal cortex describe connections between path integration and associative learning of positions. It is very likely to accelerate the understanding of spatial cognition and memory. More generally, it also provides a hypothesis and theoretical basis for human episodic memory. V. To better understand entorhinal-hippocampal neural circuit for spatial cognition and navigation, a model from grid cells in the medial entorhinal cortex to granule cells in the dentate gyrus is built to produce sparse place firing responses. It is able to solve the uncertainty caused by perception similarity during mammalian environment process, and distinct different locations in the environment. The sparse place firing responses are generated from three grid layers in the dentate gyrus granule cells by the Winner-take-all mechanism with competitive Hebbian learning rules and controlling the sparsity of dentate gyrus activity. This sparse place firing responses can break the perception of uncertainty during mammalian navigation process. This analogous algorithm, namely locality-sensitive hashing, may not only exists in the neural system of spatial cognition but also be a general principle of computation in the brain with different regions and species. VI. To theoretically explain place field expansion after inactivation of dorsal MEC, a model of Fourier spatial frequencies hypothesis based on Bayesian inference is proposed. The Fourier hypothesis is strongly supported by the proposed model on a general Bayesian mechanism to explain the expansion of place fields and predict the alignment of grid components. Furthermore, the spatial memory indexing theory is proposed to interpret the intrinsic organization of spatial navigation and episodic memory as a whole. It also describes the information transform across between neocortex (especially visual cortex and vestibular cortex), entorhinal cortex (medial entorhinal cortex and lateral entorhinal cortex), hippocampus (CA3 and CA1). It explains how new memory is formed during mammalian exploration in the entorhinal-hippocampal neural circuit and how old memory is recalled when mammals revisit the same environment for spatial cognition and navigation. In this work, the intrinsic neural mechanism of spatial cognition and navigation in the mammalian brain is investigated. In terms of the neural basis of cognitive functions, conjunctive representation of movement and space is modeled by a continuous attractor network during environment exploration; multisensory integration is realized by bayesian attractor network; an efficient compact cognitive map coding is inspired by neighborhood cells; a theoretical model is proposed to support global coherent spatial representation; a variant locality-sensitive hashing (LSH) coding is represented by place responses in the dentate gyrus able to break the perception uncertainty; place filed expansion is theoretically illuminated by Fourier grid frequency hypothesis, and spatial memory indexing theory is proposed to comprehend the framework of spatial cognition and navigation and episodic memory. In the aspect of brain-inspired intelligence, the proposed theoretical hypotheses and computational models are validated on the robotic platform and inspire novel brain-inspired mobile robot algorithms. The proposed models are able to promote better understanding the neural mechanism of spatial cognition, help build more accurate brain model to better know the biological brain; at the same time, it also provides more possible opportunities to build more intelligent, reliable, robust robot navigation system.
Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/25151
Collection机器人学研究室
Affiliation1.中国科学院沈阳自动化研究所
2.中国科学院大学
Recommended Citation
GB/T 7714
曾太平. 哺乳动物空间导航的神经计算模型研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2019.
Files in This Item:
File Name/Size DocType Version Access License
哺乳动物空间导航的神经计算模型研究.pd(36951KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[曾太平]'s Articles
Baidu academic
Similar articles in Baidu academic
[曾太平]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[曾太平]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.