READING THE NEURAL CODE IN THE MAMMALIAN VISUAL SYSTEM

研究感觉处理的一个主要挑战是理解神经元尖峰活动中加密的神经信息的含义。“阅读神经密码”的能力对于我们理解大脑功能背后的神经机制至关重要。本项目旨在了解早期视觉通路，即外侧膝状核（LGN）和初级视觉皮层中的神经编码。我们将使用实验和计算相结合的方法从两个方向来解决这个问题。在前进的方向上，我们将使用一种计算方法来表征神经元反应中编码的视觉输入的更系统的特征。相反，我们将从记录的神经元活动中重建视觉输入。在第1部分中，我们将使用解码技术从LGN中的集合响应中重建时空自然场景。在第一部分中，我们将使用解码技术从LGN中的集成响应中重建时空自然场景。在初步研究中取得成功的最优线性解码技术将用于进一步探索LGN中精确相关尖峰在视觉编码中的作用。我们还将探索使用梯度下降学习算法来训练人工神经网络以执行最佳输入重建。在第2部分中，我们将使用协方差矩阵分析来系统地表征由初级视觉皮层神经元的反应所代表的视觉输入的特征。由于大多数皮质神经元是复杂的细胞，具有高度非线性的响应，因此用常规方法很难获得对其编码特性的完整描述。协方差矩阵分析可能有助于揭示以前未知的编码特性，并可能为初级视觉皮层中神经编码的本质提供新的见解。最后，在第3部分中，我们将应用解码技术来重建自然场景，使用传统技术和协方差矩阵分析确定的属性。这将为我们在第2部分中开发的皮质视觉编码计算模型提供一个关键的测试。关于早期视觉通路的解剖学和生理学的大量信息使其成为神经编码计算分析的理想模型。这些研究的结果可能为感觉编码的一般原理提供新的见解，并可能促进我们对正常和病理条件下高级脑功能的神经元机制的理解。

A major challenge in studying sensory processing is to understand the meanings of the neural messages encrypted in the spiking activity of neurons. The ability to "read the neural code" is essential for our understanding of the neural mechanisms underlying brain functions. The proposed project aims to understand the neural code in the early visual pathway, i.e., the lateral geniculate nucleus (LGN) and the primary visual cortex. We will use a combination of experimental and computational approaches to address this problem from two directions. In the forward direction, we will use a computational method to characterize more systematically features of visual inputs that are encoded in neuronal responses. In the reverse direction, we will reconstruct visual inputs from the recorded neuronal activity. In part 1, we will use decoding techniques to reconstruct spatiotemporal natural scenes from ensemble responses in the LGN. In part 1, we will use decoding techniques to reconstruct spatiotemporal natural scenes from ensemble response in the LGN. The optimal linear decoding technique, which has been used successfully in a preliminary study, will be applied to further explore the functions of precisely correlated spiking in the LGN in visual coding. We will also explore the use of a gradient descent learning algorithm to train artificial neural networks to perform optimal input reconstruction. In part 2, we will use a covariance matrix analysis to systematically characterize the features of visual inputs that are represented by the responses of primary visual cortical neurons. Since most of the cortical neurons are complex cells with highly non-linear responses, a complete description of their coding properties is difficult to obtain with conventional methods. The covariance matrix analysis may help to reveal previously unknown coding properties and may provide new insights into the nature of the neural code in the primary visual cortex. Finally, in part 3, we will apply decoding techniques to reconstruct natural scenes using properties identified by both conventional techniques and by the covariance matrix analysis. This will provide a critical test of our computational model of cortical visual coding developed in part 2. The large amount of information on the anatomy and physiology of the early visual pathway makes it an idea model for computational analysis for neural coding. The results from the proposed studies are likely to provide new insights into the general principles of sensory coding and may advance our understanding of the neuronal mechanisms of higher brain functions under normal and pathological conditions.