Neural processing of natural scenes in the visual cortex

摘要 视网膜和视皮层以一组复杂的电信号的形式表示视觉信息。 支持视觉行为和记忆的信号。尽管我们已经学到了很多关于如何 简单的视觉模式，如条纹格栅，会导致早期视觉系统的神经活动，我们说 对自然视觉场景在行为过程中是如何表示的，以及活动的 通过身体和眼睛运动收集视觉信息的过程会影响这一过程。 视觉处理逐渐变得更加复杂，朝着大脑中更高的水平发展。相比较 对于灵长类动物，老鼠将强烈的运动输入与视觉流中较早水平的视觉输入相结合， 初级视觉皮质。这使得鼠标视觉系统能够被系统理解 自然场景如何在视觉系统中的某个层面上被主动感觉所表现和影响 其中自然场景的神经代码的计算模型更容易处理。这项建议 有两个主要目标。首先要确定自然场景的神经代码是如何变化的 视网膜到皮层的精确计算模型，可以分析以确定如何 特定的视网膜细胞类型有助于进行行为学计算的皮质活动，例如确定 运动方向和速度、自适应和对象运动检测。第二，测试替代方案 关于运动活动如何影响自然的表现的理论上的假设 场景，包括减去预期的视觉刺激以创建更有效的表示， 称为预测编码、预测或调整检测的贝叶斯特征检测 对存在视觉特征的先验概率的阈值，以及对强度的简单适应 组合信号以避免饱和。使用高通道数硅探针，计算模型 它将已知生物物理和电路级特性与可解释的尖端机器相结合 学习方法和虚拟现实系统，我们将获得对视觉处理的新见解 早期视觉系统中的自然场景。这些结果将给出视网膜如何 和视觉皮质功能，这将是理解疾病如何影响中央视觉的关键 如弱视、斜视和精神分裂症的处理，并揭示大脑皮层的一般原理 感官处理。这里建立的计算模型也将直接适用于使用 在视网膜和皮质视觉假体系统中。

Abstract The retina and visual cortex represents visual information in the form of a complex set of electrical signals to support visual behavior and memory. Although we have learned a great deal about how simple visual patterns such as striped gratings lead to neural activity in the early visual system, we know little about how natural visual scenes are represented during behavior, and how the active process of gathering visual information through body and eye movements influences this process. Visual processing becomes progressively more complex towards higher levels in the brain. Compared to primates, mice combine strong motor input with visual input at an earlier level in the visual stream, the primary visual cortex. This makes the mouse visual system an accessible to system to understand how natural scenes are represented and influenced by active sensation at a level in the visual system where computational models of the neural code for natural scenes are more tractable. This proposal has two primary goals. First to determine how the neural code changes for natural scenes from the retina to the cortex with an accurate computational model that can be analyzed to determine how specific retinal cell types contribute to cortical activity for ethological computations such as determining motion direction and speed, adaptation and object motion detection. Second, to test alternative theoretically grounded hypotheses as to how motor activity influences the representation of natural scenes, including the subtraction of expected visual stimuli to create a more efficient representation, known as predictive coding, predictive or Bayesian feature detection that adjusts the detection threshold to the prior probability that visual features are present, and simple adaptation to the strength of combined signals to avoid saturation. Using high channel count silicon probes, computational models that combine known biophysical and circuit level properties with interpretable cutting edge machine learning approaches and virtual reality systems, we will gain new insight into visual processing for natural scenes in the early visual system. These results will give a quantitative picture of how the retina and visual cortex function, which will be essential in understand how diseases that affect central visual processing such as amblyopia, strabismus and schizophrenia, and reveal general principles of cortical sensory processing. The computational models established here will also be directly applicable for use in retinal and cortical visual prosthesis systems.