CASIA OpenIR  > 毕业生  > 博士学位论文
基于脉冲神经网络的类脑情感共情与利他决策计算模型
冯慧
2024-05-17
Pages122
Subtype博士
Abstract

    情感共情是指个体感知、理解他人情感的能力,在个体生存、社会互动、群体互助协作等方面发挥着至关重要的作用。本文首先探究与总结了生物脑中情感共情以及相关利他行为决策的神经机制,并在神经元动态变化、神经递质调控以及多脑区协同的神经环路等多尺度层面对该神经机制进行计算模拟,构建了基于脉冲神经网络的类脑情感共情与利他决策计算模型。同时,本文针对该模型中不同的子功能,分别设计了对应的智能体行为任务实验,以验证模型的功能层面有效性,致力于使智能体具备机制类脑、行为类生物的情感共情以及利他能力。

    在生物脑中,情感共情能力是通过镜像神经系统的镜像机制实现的。个体在体验自身情感并产生表情、声音等情感外显动作过程中,通过感觉运动经验训练形成相应的镜像神经元。在感知到他人同样的情感外显动作信息时,会经由镜像神经元介导,激活自身对应的情感神经表征,通过具身体验实现对他人情感的理解。同时,共情到的情感结果还会通过影响自身的奖励系统,改变多巴胺等神经递质的释放浓度,为利他行为提供内在动机。基于上述机制,类脑情感共情与利他决策计算模型分为三个子模型,分别负责对情感产生的神经机制、基于镜像神经系统的情感共情神经机制以及由情感共情驱动的利他决策神经机制进行计算建模。具体工作以及创新点归纳如下:

    一、类脑机器“疼痛”脉冲神经网络模型。国际疼痛协会定义疼痛为一种与实际或潜在的组织损伤相关或类似的不愉快的感觉和情感体验,对生物体的生存具有重要意义。同时,疼痛能够产生相应的情感外显信息,如痛苦的表情、喊叫声等,可作为引发共情以及利他行为的先决条件。本研究以疼痛这种特殊的情感作为研究对象,首先探究生物脑中由实际和潜在身体损伤引发疼痛体验的相关脑区以及神经环路,并受其启发使用脉冲神经网络以及脉冲时序依赖可塑性学习法则对前扣带回皮层、腹侧被盖区等多个脑区的功能及其连接进行计算模拟,构建了一个类脑机器“疼痛”脉冲神经网络模型,其具备高度的生物合理性。本工作还设计了机器人的实际机体损伤实验以及潜在机体损伤实验,以验证模型的有效性。实验结果表明,该模型通过表征和关联机器人多模态感觉信息、运动信息与身体状态之间的耦合关系,成功实现了对于实际机体损伤事件的辨别和报警,产生类生物的机器“疼痛”状态。同时,该机器“疼痛”状态还通过关联学习建立了与损伤相关触发线索之间的联系,使机器人能够对潜在损伤产生快速反应,并通过影响多巴胺信号的强弱完成对潜在危险的回避学习,体现了与生物体疼痛一致的功能和意义。

    二、类脑情感共情脉冲神经网络模型。本工作以类脑机器“疼痛”脉冲神经网络计算得出的机器“疼痛”结果作为初始负面情感输入,并对生物脑中基于镜像神经系统的情感共情神经机制进行计算建模,构建了一个类脑情感共情脉冲神经网络模型。该模型计算模拟了大脑中参与情感共情过程的情感脑区、运动脑区、感知脑区以及这些脑区之间的连接。在体验自身情感以及产生情感外显动作时,该模型根据感觉运动经验,使用脉冲时序依赖可塑性学习法则训练形成相应情感外显动作的镜像神经元和反镜像神经元,并作用于情感共情过程。本工作设计了双智能体情感共情任务实验进行模型的功能验证,结果表明,在感知他人情感外显动作信息时,该模型能够成功利用镜像神经元激活自身表征同一情感的情感神经元,实现对同伴的情感共情,并利用反镜像神经元进行自我-他人区分。此外,受共情的个体差异性启发,该计算模型通过在输入端设计不同的抑制性突触连接比例成功建模了不同的情感共情水平,并得到了与神经科学研究结论一致的实验结果。

    三、情感共情驱动的类脑利他决策脉冲神经网络模型。本工作探究和总结了负面情感共情驱动的利他行为决策的神经机制,受其启发构建了一个类脑利他决策脉冲神经网络模型。该模型计算模拟了包括腹侧被盖区、基底神经节在内的多个决策相关脑区的功能,以类脑情感共情脉冲神经网络计算得到的共情结果作为输入,并对不同共情结果对应的内在奖励值进行计算。该模型使用奖励调控的脉冲时序依赖可塑性学习法则来训练负责动作选择的突触权重,从而指导利他行为决策。本工作设计了一个双智能体利他救援任务实验,以验证模型的利他决策功能。实验结果表明,该模型能够成功兼顾利他任务和自身任务,表现出明显的利他性。此外,该模型还通过实验探究了不同共情水平对利他行为表现的影响,结果表明共情水平与利他倾向呈正相关,共情水平高的智能体具备更强烈的利他倾向,愿意牺牲更多的自身利益执行利他行为。该结果与行为学研究中的利他决策实验结果相一致,进一步说明了所提模型的有效性和合理性。

    本文采用递进式的研究思路,分别对生物脑中疼痛产生、情感共情以及相关利他决策的神经机制进行计算模拟,构建了一个相对完整的类脑情感共情与利他决策计算模型,并通过相应的实验任务以及与神经科学实验数据的比较分析验证了所提模型的功能有效性以及生物合理性。本文不仅为计算领域构建具备情感理解以及自主利他能力的智能系统提供了新的研究思路,也为神经科学领域进一步揭示情感共情以及相关行为的神经机制提供了计算基础。

Other Abstract

Affective empathy refers to an organisms' ability to perceive and understand other's emotions, which plays a crucial role in organism survival, social interaction, and group collaboration. This thesis explores and summarizes the neural mechanisms of affective empathy and related altruistic decision-making in the biological brain, and computationally simulates these neural mechanisms at the multi-scale levels of neuronal dynamics, neurotransmitter regulation, and neural loops of multi-brain regions, and constructs a brain-inspired affective empathy and altruistic decision computational model based on the spiking neural networks. Meanwhile, this thesis designs corresponding intelligent agent behavioral task experiments for different sub-functions to verify the effectiveness of the model at the functional level. This thesis aims to equip intelligent agents to possess the affective empathy and altruistic ability of mechanism-like brain and behavior-like organisms.

In the biological brain, affective empathy is realized through the mirror mechanism of the Mirror Neuron System. Individuals form corresponding mirror neurons through sensory-motor experience training during the process of experiencing their own emotions and generating emotional overt actions such as expressions and sounds. When they perceive the same emotional overt action information from others, they will activate the corresponding emotional neural representations through mirror neurons, and realize the understanding of others' emotions through embodied experience. At the same time, the empathic emotion can provide intrinsic motivation for individuals' altruistic behaviors by influencing  reward system and altering the concentration of dopamine or other neurotransmitters. Based on the above mechanisms, the brain-inspired affective empathy and altruistic decision computational model is divided into three sub-models, which are respectively responsible for computational modeling of the neural mechanism of emotion generation, the neural mechanism of affective empathy based on the Mirror Neuron System, and the neural mechanism of altruistic decision-making driven by affective empathy. The specific work and innovations are summarized below:

1)Brain-inspired Robot Pain Spiking Neural Network. The International Association for the Study of Pain defines pain as an unpleasant sensory and emotional experience associated with, or resembling that associated with actual or potential tissue damage, which is important for the survival of organisms. At the same time, pain can generate corresponding emotional overt action, such as painful expressions and shouts, which are equipped to trigger pain empathy as well as altruistic behavior. This thesis explores the neural mechanism involved in pain experience caused by actual and potential physical injuries in the biological brain, and inspired by this mechanism, constructs a brain-inspired robot pain spiking neural network. This model uses spikng neurons as well as the Spike-Timing-Dependent Plasticity learning rule to computationally simulate the functions and connectivity of several pain-ralated brain regions, including the Anterior Cingulate Cortex and the Ventral Tegmental Area. This thesis designs an actual body injury experiment and a potential body injury experiment of the robot to verify the validity of the proposed model. The experimental results show that the proposed model successfully realizes the recognition and alarm of actual robot body injury events by characterizing and correlating the coupling relationship between multimodal sensory information, motor information and body states of the robot, and generates a biologically similar Robot Pain state. In addition, this Robot Pain state establishes a connection to injury-related trigger cues through associative learning, enabling the robot to quickly react to subsequent relevant cues and accomplish avoidance learning for potential dangers through the modulation of dopamine signaling. Therefore, the Robot Pain constructed by the proposed model is functionally and meaningfully consistent with biological pain.

2)Brain-inspired Affective Empathy Spiking Neural Network. Inspired by the neural mechanism of affective empathy based on Mirror Neuron System in the biological brain, this thesis constructs a brain-inspired affective empathy spiking neural network. This model computationally simulates the emotional brain regions, motor brain regions, perceptual brain regions, and the connections between them involved in the affective empathy process in the biological brain, and receives the Robot Pain result computed by the brain-inspired robot pain spiking neural network as the initial negative emotional input for the training of affective empathy ability. When experiencing self emotion and generating emotional overt action, the proposed model uses the Spike-Timing-Dependent Plasticity learning rule to train the mirror neurons and anti-mirror neurons corresponding the emotional overt action based on sensory-motor experiences to act on the empathy process. This thesis designs a two-agent affective empathy task experiment for the functional validation of the proposed model, and the results show that when perceiving the information of peer's emotional overt action, the agent applying the proposed model can use mirror neurons to activate its own emotional neurons that characterize the same emotion to successfully achieve empathy for its peer, and use anti-mirror neurons to differentiate between self and other. In addition, inspired by the individual variability of empathy, the proposed model simulates different levels of affective empathy by designing different proportions of inhibitory synaptic connections, and obtains experimental results consistent with neuroscience findings.


3)Brain-inspired Altruistic Decison-making Spiking Neural Network Driven by Affective Empathy. This thesis explores and summarizes the neural mechanisms of altruistic behavior decision driven by negative emotional empathy, and inspired by these mechanisms, constructs a brain-inspired altruistic decision-making spiking neural network. This model simulates the functions of multiple decision-related brain regions including the Ventral Tegmental Area and the Basal Ganglia, and receives the empathic outcomes computed by the brain-inspired affective empathy spiking neural network as inputs. This model calculates the intrinsic reward values corresponding to different empathy outcome, and uses a reward-modulated Spike-Timing-Dependent Plasticity learning rule to train the synaptic weights responsible for action selection to guide altruistic decision-making. This thesis designs a two-agent altruistic rescue task experiment to verify the proposed model's altruistic decision-making performance. The experimental results show that the proposed model is able to successfully balance the self-task and the altruistic-task, demonstrating obvious altruism. In addition, the proposed model also experimentally verifies the effects of different empathy levels on altruistic behavior performance, and the result shows that the empathy levels are positively correlated with altruistic tendencies. The agents with high empathy levels possess stronger altruistic tendency and are willing to sacrifice more self-interests to perform altruistic behaviors. This result is consistent with the experimental result of altruistic decision-making in behaviorism research, further proving the validity and rationality of the proposed model.

This thesis progressively simulates the neural mechanisms of pain generation, affective empathy, and altruistic decision-making in the biological brain, and constructs a relatively complete a brain-inspired affective empathy and altruistic decision computational model, and verifies the functional validity and biological rationality of the proposed model through the corresponding experimental tasks and comparative analyses with neuroscience experimental data. This thesis not only provides a new research idea for the construction of intelligent systems with emotional understanding and autonomous altruism in the artificial intelligence field, but also provides a computational basis for the neuroscience field to further reveal the neural mechanisms of affective empathy and related behaviors.

Keyword情感共情,利他决策,脉冲神经网络,突触可塑性,多脑区协同
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/57359
Collection毕业生_博士学位论文
毕业生
Recommended Citation
GB/T 7714
冯慧. 基于脉冲神经网络的类脑情感共情与利他决策计算模型[D],2024.
Files in This Item:
File Name/Size DocType Version Access License
基于脉冲神经网络的类脑情感共情与利他决策(18253KB)学位论文 限制开放CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[冯慧]'s Articles
Baidu academic
Similar articles in Baidu academic
[冯慧]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[冯慧]'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.