毕业生知识库

      本文基于当前神经科学领域关于符号表征与推理的研究成果，构建了多环 路协同的类脑脉冲神经网络模型，系统性地针对脉冲神经网络如何完成符号表 征以及推理任务展开讨论与研究。从模型功能和结构上看，本文从微观尺度，介 观尺度和宏观尺度分别对大脑的连接结构和功能进行了借鉴。在微观尺度上，本 文采用了类生物神经元的计算模型，引入脉冲时序依赖可塑性机制 (STDP)，以 及奖励信号调制的脉冲时序依赖可塑性机制 (R-STDP) 来建立网络中神经元间的 微观突触结构。在介观尺度上，本文受神经元群体编码机制启发，以神经元组完 成对于概念符号的表征，并通过学习与记忆使得不同神经元组间形成稳定的突 触连接，构建了多环路协同类脑脉冲神经网络实现符号序列表征与重建任务。宏 观功能层面，本文详细阐述了基于数据驱动的深度神经网络缺乏因果推理能力， 与人类的智能存在本质上的差异，并在宏观功能层面探索验证了类脑脉冲神经 网络完成因果推理的可行性。
      本文的主要工作和创新点归纳如下:
      第一，基于群体编码的符号序列生成类脑脉冲神经网络模型。受神经科学领 域对于猕猴如何完成符号表征与符号序列生成的相关研究启发，本文提出了一 种多环路协同的类脑符号序列生成脉冲神经网络来模拟与猕猴相同的符号序列 记忆与生成过程。经过实验验证，证明了通过融合群体编码表征机制和脉冲时序 依赖可塑性机制，网络可以完成对于符号序列的记忆。通过引入基于奖励信号调 制的脉冲时序依赖可塑性机制，经强化学习后，本网络可以按照不同的超正则语 法规则实现对于符号序列的重建。并且，网络在实验中的表现与猕猴的行为学数 据可以很好拟合，该发现进一步展示了网络的生物合理性。此外，本工作还验证 了组块 (Chunking) 机制在提高模型的鲁棒性方面起到了重要作用。
      第二，类脑常识知识表征与推理脉冲神经网络。受生物脑中神经元群体编 码符号概念的神经机制启发，本文融合了群体编码表征机制和脉冲时序依赖可 塑性学习法则，将常识知识库中的常识知识编码进大规模的脉冲神经网络当中， 并据此完成了后续相关推理任务。各个神经元组通过突触连接共同编码了整个 常识知识图谱，形成了一张记录常识知识的图脉冲神经网络。此外，引入了奖励调制的脉冲时序依赖可塑性机制来实现类生物的自监督强化学习过程，完成了 相关的推理任务。针对相同的推理任务，与图卷积人工神经网络相比，本工作获 得了可比较的准确率和更快的收敛速度。
      第三，基于因果图的类脑因果推理脉冲神经网络。当前基于数据驱动的深度 神经网络方法本质上只利用了数据之间的相关性，难以透过数据看到世界的因 果本质。赋予机器因果推理能力是迈向更高水平通用智能的关键一步。为解决这 一问题，受基于因果图的推理方法启发，结合大脑中相关神经机制和连接结构， 我们提出了类脑因果推理脉冲神经网络模型。本模型利用群体编码机制和脉冲 时序依赖可塑性学习法则完成了将静态因果图编码进脉冲神经网络当中。通过 实验验证，证明了脉冲神经网络完成因果推理的可行性。并且，通过引入外部评 测函数，观测者可以根据网络的放电状态得知网络在完成推理过程中的推理路 径，进而大幅提升了网络的可解释性，为进一步探索构建人类可理解的脉冲神经 网络系统打下基础。

      Based on current research results in neuroscience on symbolic representation and reasoning, this paper constructs a multi-loop collaborative brain-inspired spiking neural network model, and systematically discusses and investigates how spiking neural networks accomplish symbolic representation and reasoning tasks. In terms of model function and structure, this paper draws on the brain's function and connectivity structure at the microscopic, mesoscopic and macroscopic scales respectively. At the microscopic scale, the paper adopts a computational model of biological neuron-like neurons and introduces spiking timing-dependent plasticity mechanisms for synapses as well as reward-modulated spike timing-dependent plasticity mechanisms to build the micro-structure of neural networks. At the mesoscopic scale, inspired by the population coding mechanism, this paper uses neuron populations to complete the representation of symbols, and through learning and memory to form stable synaptic connections between different neuronal populations, a multi-loop synergistic brain-inspired spiking neural network is constructed to achieve the task of symbol sequence representation and reconstruction. At the macro level, the paper elaborates that data-driven deep neural networks lack causal reasoning and are fundamentally different from human intelligence, and explores the feasibility of brain-inspired spiking neural networks for causal reasoning at the macro functional level.

      The main work and innovations of this paper are summarized as follows:

      First,  Brain-inspired Sequence Production Spiking Neural Network (SP-SNN). Inspired by the related research on how macaques complete symbolic representation and symbolic sequence production in neuroscience, this paper proposes a multi-loop coordinated brain-inspired symbolic sequence production spiking neural network to simulate the same symbolic sequence memory and production process as in macaques. After experimental verification, it is proved that SP-SNN can complete the working memory of the symbol sequence by fusing the population coding representation mechanism and the spiking timing-dependent plasticity mechanism (STDP). Moreover, by introducing the reward-modulated spike timing-dependent plasticity (R-STDP) mechanism based on the modulation of the reward signal, after reinforcement learning, SP-SNN can produce the symbol sequence according to different supra-regular grammar rules. Moreover, SP-SNN's performance in the experiment can be well fitted with the behavioral data of macaques, which further demonstrates the biological plausibility of the network. In addition, this work also verifies that the chunking mechanism plays an essential role in improving the robustness of the model.

      Second, Brain-inspired Graph Spiking Neural Networks for Commonsense Knowledge Representation and Reasoning (KRR-GSNN). Inspired by the neural mechanism of the neuron population coding symbolic concept in the biological brain, KRR-GSNN combines the population coding representation mechanism and the spike-timing-dependent plasticity (STDP) learning rule to encode the commonsense knowledge in the commonsense knowledge graph into a large-scale spike neural network and completed subsequent related reasoning tasks accordingly. Each neuron population jointly encodes the entire commonsense knowledge graph through synaptic connections, forming a graph spiking neural network that records commonsense knowledge. Furthermore, the reward-modulated spike-timing-dependent plasticity mechanism (R-STDP) is introduced to realize the biological-inspired self-supervised reinforcement learning process and complete related reasoning tasks. KRR-GSNN achieves comparable accuracy and faster convergence for the same reasoning task than graph convolutional artificial neural networks.

      Third, Brain-Inspired Causal Reasoning Spiking Neural Networks (CR-SNN). The current data-driven deep neural network methods essentially only utilize the correlation between data, difficult to see the causal nature of the world through data. Empowering machines with causal reasoning ability is a critical step toward a higher level of general intelligence. Inspired by the reasoning method based on the causal graph, we propose brain-inspired causal reasoning spiking neural network model by combining the relevant neural mechanisms and connection structures to solve this problem. CR-SNN completes the construction of a static causal graph by using the population coding mechanism and the learning rule of spiking timing-dependent plasticity. The feasibility of causal reasoning by spiking neural networks (SNN) is proved through experimental verification. Moreover, by introducing an external evaluation function, observers can know the reasoning path of the network in the completion of the reasoning process according to the firing state of the network, thereby significantly improving the interpretability of the network, laying a foundation for further exploration and construction of human-comprehensible spiking neural network systems.