RUI: Computational Modeling of High Level Cortical Motion Processing

一个人在世界上移动时，必须判断自己的运动路径和场景的三维布局，并识别场景中的运动物体。例如，为了到达期望的目的地并避免碰撞，在忙碌街道上的驾驶员必须能够将汽车保持在其车道上，判断其他汽车和行人的运动，并判断场景中的其他对象位于何处。所有这些判断都可以根据聚焦在视网膜上的视觉图像的运动来做出。 实验表明，人们非常擅长完成这些任务，这意味着必须有神经机制来计算观察者的运动，场景布局和移动物体的存在。 在这组研究中，将开发这种神经处理的计算模型，以研究计算这些运动特性的潜在神经过程。 该模型将基于当前对视觉皮层中颞区（MT）中的单个神经元如何响应视觉运动的知识。 已知MT区域有许多对运动反应良好的神经元。 建立这样一个计算模型的目的是双重的。 首先，它将展示视觉皮层中的神经元如何计算观察者运动的方向和表面的相对深度，以及它们如何在观察者运动期间识别和定位场景中的移动物体。 此外，模型应该生成关于神经元预期行为的清晰、可测试的预测，这些神经元使用图像中的运动来计算有关场景的信息。 这些预测将导致更多的实验，进一步扩展我们对这些神经元的认识。 计算模型将通过运行模拟来评估，这些模拟将检查其计算观察者运动方向、场景布局和移动物体位置的能力。 模拟的结果将显示模型是否表现得与人类观察者一样好，以及模型的输出是否复制了实验测试期间人类观察者的行为。 进一步的测试将检查模型中的计算单元是否与视觉皮层的运动处理区域中的细胞表现相同，例如MT和内侧上级颞区，这部分视觉皮层被认为参与计算观察者运动的方向。这种视觉运动神经处理模型的实现和测试将有助于加深对大脑如何分析视觉世界的理解。

A person moving through the world must judge his or her own path of motion and the three dimensional layout of the scene and identify moving objects in the scene. For example, to reach a desired destination and avoid collisions, a driver on a busy street must be able to keep the car in its lane, judge the motion of other cars and pedestrians and judge where other objects in the scene are located. All of these judgments can be made based on the motion of the visual images that are focused on the retina. Experiments show that people are remarkably adept at accomplishing these tasks, which implies that there must be neural mechanisms that compute observer motion, scene layout and the presence of moving objects. In this set of studies a computational model of this neural processing will be developed in order to investigate the underlying neural processes that compute these motion properties. The model will be based on current knowledge of how individual neurons in the Middle Temporal area (MT) of the visual cortex respond to visual motion. The MT area is known to have many neurons that respond well to motion. The purpose of building such a computational model is two-fold. First, it will show how neurons in visual cortex may calculate the direction of observer motion and the relative depth of surfaces and how they identify and locate moving objects in the scene during observer motion. In addition, the model should generate clear, testable predictions about the expected behavior of neurons that use motion in the image to compute information about the scene. These predictions will lead to additional experiments that will further extend our knowledge of these neurons. The computational model will be evaluated by running simulations that examine its ability to compute the direction of observer motion, the scene layout and the location of moving objects. The results of the simulations will show whether the model performs as well as human observers and whether the output of the model replicates the behavior of human observers during experimental testing. Further tests will examine whether the computational units in the model behave in the same way as cells in the motion processing areas of the visual cortex, such as MT and the Medial Superior Temporal area, the part of visual cortex thought to be involved in computing the direction of observer motion. The implementation and testing of this model of neural processing of visual motion will lead to increased understanding of how the brain analyzes the visual world.