感知、协作与进化—智能依赖于结构和其上的运动

发布时间：2017-12-27 20:17

  本文关键词：感知、协作与进化—智能依赖于结构和其上的运动　出处：《南京大学》2016年博士论文　论文类型：学位论文

  更多相关文章： 感知协作神经网络 感知进化神经网络 人工进化 皮质-感受器人工扩展

【摘要】：受生物学启发的计算模型是促进人工智能领域发展的重要推动力。长期以来,大量的学者根据生物现象提出了各种各样的仿生计算模型,为人工智能领域做出了许多的开创性的贡献。本文试图跟随这样的足迹,调研了近些年一些生理学实验及发现,并在此基础上发现,神经系统的结构以及神经冲动在神经系统结构上的运动模式对于生物体的智能行为有着至关重要的影响。可以说,生物体的智能行为依赖于神经系统的结构和神经冲动在该结构上的运动模式。在这种想法下,本文试图模拟大脑的结构、神经冲动在该结构上的运动模式、以及这种结构的开放性(或可扩展性),提出了两种仿生计算模型,即感知协作神经网络和感知进化神经网络。主要工作如下：(1)为了模仿大脑对于不同感官感觉的协作,提出了感知协作神经网络。该网络的建立受到大脑的层级结构以及层级结构中各部分功能单元相互协作的启发。感知协作神经网络为层级结构,按照功能对应于初级感觉区、初级感觉联合区和高级联合区。初级感觉区负责处理一些零碎的感觉,如颜色、形状和音节等。初级感觉联合区连接初级感觉区中零碎的感觉,形成表征单一物体的概念,如视觉实体概念、听觉概念和味觉实体概念等。高级联合区连接不同的初级感觉联合区,如视听感觉、视味感觉的联合等,即该区域执行通感功能。感知协作神经网络对神经元进行了功能职责的划分,包括特征神经元、初级概念神经元和联合神经元。不同类别的神经元拥有不同的行为模式。感知进化神经网络可以通过建立神经元之间的突触连接而快速形成新的记忆。无意识冲动和内省机制用以保证网络结构与外部数据结构的一致性。该模型可以应用于概念获取、信息融合、在线学习系统、机器人系统等领域。(2)文献[87]中的实验表明,cDNA基因敲入小鼠表现出增强的长波长光感知能力并获得新的颜色辨别能力。该实验暗示了生物体的感知系统或神经系统结构是可以人工扩展的,如利用[87]文中的基因工程技术。受到上述启发,本文提出如下问题：是否可以开发一个可以在线扩展感知能力的智能主体,从而打破智能主体的感知限制。为解决这一问题,提出了感知进化神经网络,包括类型I，新型感受器加入已有感官和类型II，新型感官突现于智能体。当新型感受器加入已有感官时,新型感受器接受的刺激连同已有感受器接受的刺激传入到初级感觉区,此区域中的特征神经元把从固有感受器提取的特征和从新型感受器提取的特征进行联合,从而形成更加深化的特征概念。即通过在特征神经元和新型感受器之间建立突触,使特征神经元的响应维度被在线扩充。当新型感官突现于智能体时,新型感官感受器接受的刺激连同已有感官感受器接受的刺激传入到初级感觉区并在此区域加工成零碎的感觉。这些零碎的感觉沿着网络结构向上传导至初级感觉联合区,并在该区域激活响应的概念神经元。而后,概念神经元将激活信号传导至高级联合区,新型感官与已有感官将在此区域通过联合神经元进行联系,从而形成新型感官和固有感官的协作。该模型从计算的观点出发,部分回答了[87]文中关于“…创建额外类型的感觉神经元或者促进神经回路的突现,以比较新的和现有的感觉反应”的问题。感知进化神经网络可以应用于信息融合、在线学习系统、机器人系统等领域。(3)种种神经生理学实验[142,146,187]和医学案例迹象表明大脑皮质和生物体感受器是具有可扩展能力的。基于感知协作神经网络、感知进化神经网络和这些神经生理学实验以及医学案例,本文提出了皮质-感受器人工扩展理论。基于该理论进而提出人工进化的概念,即人工进化旨在通过感官-大脑重整和扩展在生命体级别上实现进化。虽然目前该概念和理论处于设想阶段,但是这一概念和理论具有很好的可挖掘性,也同时具有很好的启示性。感知协作神经网络和感知进化神经网络旨在模拟生物的感知、协作以及进化行为。文中实验初步验证了这两种模型的有效性。然而,想要将这两种模型投入到实际应用中去,还有很长的路要走。论文的后语部分讨论了一些问题,主要集中在微观的生理结构和生理过程是如何和宏观的智能行为进行联系。文中对于这些问题的讨论虽然比较粗糙,但是这些都是值得深入思考的问题。希望这些问题可以有后续的理论化和工程化。
[Abstract]:The computational model inspired by biology is an important driving force in the development of artificial intelligence. For a long time, a large number of scholars have proposed a variety of bionic computing models based on biological phenomena, which have made many pioneering contributions to the field of artificial intelligence. This paper attempts to follow such a footprint and investigate some physiological experiments and findings in recent years. On this basis, it is found that the structure of the nervous system and the movement mode of nervous impulse on the nervous system structure have a crucial impact on the intelligent behavior of organism. It can be said that the intelligent behavior of the organism depends on the structure of the nervous system and the pattern of the movement of the nerve impulses on the structure. Under this idea, we try to simulate the brain's structure, the movement mode of nerve impulse on this structure, and the openness (or extensibility) of the structure. We propose two bionic computing models, namely the perceptive cooperative neural network and the perceptive evolutionary neural network. The main work is as follows: (1) in order to imitate the cooperation between the brain and different sensory senses, a cognitive cooperative neural network is proposed. The establishment of the network is inspired by the hierarchical structure of the brain and the collaboration of functional units in the hierarchical structure. The cognitive cooperative neural network is a hierarchical structure that corresponds to the primary sensory area, the primary sensory Union and the advanced United region according to its function. The primary sensory area is responsible for dealing with some fragmentary feelings, such as color, shape, and syllable. The primary sensory association area connects the fragmented sensation in the primary sensory area to form the concept of representing a single object, such as the concept of visual entity, auditory concept and taste entity concept. The high level united region connects different primary sensory associations, such as audio-visual sensation, and the union of visual taste, that is, the region performs synaesthesia. The cognitive cooperative neural network is divided into functional functions of neurons, including characteristic neurons, primary conceptual neurons and joint neurons. Different types of neurons have different behavioral patterns. Cognitive evolutionary neural networks can quickly form new memories by establishing synaptic connections between neurons. The unconsciousness and introspection mechanism are used to ensure the consistency of the network structure and the external data structure. The model can be applied to the fields of concept acquisition, information fusion, online learning system, robot system and so on. (2) showed that in [87] experiment, cDNA gene knock in mice showed long wavelength perception enhancement and obtain new color discrimination ability. The experiment suggests that the perceptual system of the organism or the structure of the nervous system can be expanded artificially, such as the use of genetic engineering in [87]. Inspired by the above, this paper puts forward the following question: can we develop an intelligent agent that can expand the perception ability online, thereby breaking the perception limit of the intelligent subject. To solve this problem, we propose a perceptive evolutionary neural network, including type I, new sensory receptors, adding existing senses and types of II, and new senses emerging into agents. When new sensors join the existing senses when new receptors receive stimulation with existing receptor for the stimulation of afferent to the primary sensory areas, characteristics of neurons in this area from the feature extraction and the inherent feeling is extracted from the new sensor characteristics are combined, formed from the concept of deepen. By establishing synapses between the characteristic neurons and the new-type receptors, the response dimensions of the characteristic neurons are extended online. When the new sensory organ emerges in the intelligent body, the stimulation received by the new sensory receptor is introduced into the primary sensory area with the stimulus received by the sensorimotor, and processed into fragmented sensation in this area. These fragmentary sensations are transmitted upwards along the network structure to the primary sensory Union and activate the conceptual neurons in response in the region. Then, the concept neuron will activate the signal transduction to the advanced joint area, and the new sense organ and the existing sense will connect with the combined neuron in this area, so as to form new sensory and inherent sensory cooperation. From the point of view of the calculation, the model answers part of the [87] article about "..." Create additional types of sensory neurons or promote the emergence of neural circuits to compare the problems of new and existing sensory responses. The cognitive evolutionary neural network can be applied to information fusion, online learning system, robot system and other fields. (3) a variety of neurophysiological experiments [142146187] and medical cases show that the cerebral cortex and biological receptors are extensible. Based on perceptive cooperative neural network, perceptive evolutionary neural network, and these neurophysiological experiments and medical cases, the theory of cortical receptor artificial expansion is proposed in this paper. Based on this theory, the concept of artificial evolution is proposed, that is, artificial evolution aims to achieve evolution at the life level through the reorganization and expansion of the senses - the brain. Although the concept and theory are at the tentative stage at present, the concept and theory are very good to be excavated and have good inspiration. Cognitive cooperative neural networks and perceptual evolutionary neural networks are designed to simulate the perception, collaboration and evolutionary behavior of organisms. The validity of the two models is preliminarily verified in the experiment. However, there is a long way to go to put these two models into practical applications. The latter part of the thesis discusses some problems, mainly focusing on how the microscopic physiological structure and physiological processes relate to the macro intelligent behavior. Although the discussion of these problems is relatively rough, these are all problems worth thinking deeply. It is hoped that these problems can be further theorized and engineered.
【学位授予单位】：南京大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP18
，

本文编号：1343047