整合不同空间位置上的视觉信息是大脑视觉系统的重要功能之一，外周抑制作用作为反映神经元整合能力的最重要指标之一，在包括啮齿类和灵长类等不同物种的多个视觉相关脑区中普遍存在。然而，即使是分层结构和功能特性被研究得最为清楚的初级视皮层(primary visual cortex, V1)，神经元外周抑制的计算原理目前仍不清楚。尤其是在V1的输出层，神经元接收来自于前馈及其他环路（包括层内的局部连接、长程的水平连接以及高级皮层的反馈连接）的协同调制时，多种环路通过何种计算方式影响神经元的空间整合特性是尚未解决的问题。

为了区分不同作用机制对V1外周抑制的影响和贡献，探究V1神经元整合大范围视觉信息的计算原理，本研究使用多通道线性电极同时记录清醒猕猴V1不同亚层神经元对不同尺寸和类型视觉刺激的电信号，结合神经元反应特性的刻画和模型解析的方法进行了以下研究：

在研究一中，我们精确刻画了V1各亚层神经元进行大范围视觉信息整合的反应特性。本研究揭示了V1神经元进行大范围视觉信息整合时以抑制为主要表现形式的跨层变化规律：输出层神经元有比输入层更强的外周抑制，同时我们首次发现猕猴V1的输出层神经元，相比于输入层神经元有更小尺寸的经典感受野以及独特的中心抑制现象。

在研究二中，我们关注形成V1各亚层神经元空间整合特性的计算原理。首先使用已有的经典模型同时捕捉V1各亚层神经元对不同尺寸圆盘和空环刺激的反应特性时，我们发现虽然除法形式的计算模型可以解释输入层神经元的整合特性，但是单一的除法或减法模型均无法解释输出层神经元的整合特性。为了解释输出层神经元的反应特性，我们首次提出了级联除法模型，并将该模型扩展到二维形式。结果表明，无论是一维还是二维形式的级联除法模型均可以很好地解释输出层神经元的反应特性。此外，我们的结果也证明了级联除法模型具有较好的泛化能力。

在研究三中，我们解析V1各亚层神经元所应用的计算原理在整合大范围空间信息中的作用及可能的环路来源。我们通过对级联除法模型构成要素的拆分，解析了不同计算方式对神经元空间整合特性的影响方面。结果表明除法成分对V1神经元的多个整合特性有重要作用。此外，我们发现级联除法模型中的除法成分具有更广的空间范围，推测其可能源于长程的水平连接或者高级皮层的反馈连接；而减法成分具有较小的空间范围，推测其可能源于层内的局部连接。

在研究四中，基于我们揭示的V1神经元整合大范围视觉信息的反应规律和计算原理，我们衡量了多个卷积神经网络(convolutional neural network, CNN) 对大范围空间信息的整合特性及其与V1的异同，并且探究了CNNs中归一化操作（除法）对人工神经元空间整合特性的影响。我们发现CNNs中底层人工神经元的整合特性与V1神经元有显著的差异，同时CNNs中的除法作用也与V1不同。

综上所述，我们刻画了V1各层神经元进行大范围视觉信息整合的反应特性并揭示了其背后的计算原理。同时，我们发现CNNs中的人工神经元和生物脑中的V1神经元在空间整合特性和所应用的计算原理中均存在显著的差异。本研究不仅加深了我们对灵长类V1区整合视觉信息工作原理的认识，同时也为优化和改善人工神经网络提供了新的思路和启发。

Spatial integration is one of the important functions of visual processing in the brain. Surround suppression, as one of the most critical indicators of spatial integration, exists throughout multiple visual brain areas across species. But even in the primary visual cortex (V1), the computational principle for spatial integration is still unclear, especially for those neurons in the output layers of V1. Neurons utilize what computations to fulfill spatial integration under feedforward connections cooperating with other circuits, including local recurrent connections,long-range horizontal connections, and feedback connections from the higher-level cortex are the unresolved issue.

To reveal the functions of multiple circuitry connections and to explore the computational principles of spatial integration in V1, we simultaneously recorded neural responses in different layers by the multiple-channel linear array. Combined with characterizations of the spatial integration's properties of the neurons and model analysis across different layers, the following research is carried out:

In the first part of the thesis, we characterized the response properties of neurons across different layers in V1 for spatial integration. We found three significant laminar differences in the spatial integration within V1, which are the stronger surround suppression, smaller receptive field size, and specific center suppression for neurons inthe output layers compared to neurons in input layers.

In the second part of the thesis, we focused on the computations for spatial integration in V1. We first captured the suppressive effect of neurons in different layers using two classical computations, namely subtraction and division. The results showed that neither a simple subtraction nor division could explain the spatial integration of neurons in the output layers; a simple division could explain the properties in the inputlayers. Then, we offered a new cascaded normalization model (CN model) to explainthe spatial integration of neurons in the output layers. Both one-dimensional and two-dimensional models showed good fitting performances. Further, the CN model also showed good generalizability.

In the third part of the thesis, we analyzed the contributions and potential circuitry mechanisms for the different computations modulating the spatial integration in V1. We analyzed what factors the different computations mainly affect the spatial integration bydisassembling the model components. We found that the divisive components have an important impact on the spatial integration of V1 neurons. In addition, we found that the divisive component occupied a global spatial extent and may originate from long-range horizontal connections or feedback connections in the higher visual cortex; the subtractive component occupied a local spatial extent and may originate from local recurrent connections within the layers.

In the fourth part of the thesis, based on our results of the computational principles for spatial integration in V1, we compared the similarities and differences between the convolutional neural networks (CNNs) and V1. We found that the spatial integration of artificial neurons and the functions of division in CNNs were different fromthose of neurons in V1.

In summary, we characterized the spatial integration of neurons across different layers of V1 and revealed the underlying computations for spatial integration. However, there are significant differences between CNNs and V1. Based on our research, it not only deepens our understanding of how the visual cortex integrates visual information but also provides new ideas for improving CNNs.