Distributed neural processing of self-generated visual input in a vertebrate brain

我们在日常生活中所做的大部分事情都涉及到我们的感官和行动之间持续而流畅的协调。例如，泡一杯茶需要处理持续不断的视觉和触觉信息流，以不断纠正我们运动肌肉的方式，以避免洒出牛奶或打碎杯子。我们的大脑可以毫不费力地协调这种从感觉到运动的信息流，但这是现代工程仍然无法匹敌的能力；想想今年DARPA机器人挑战赛上笨拙的机器人吧。在帕金森氏症和运动障碍等疾病中，协调感觉输入和运动动作的能力也经常受损。因此，理解这种能力不仅是现代神经科学的中心目标，也是推动工程进步和促进疾病治疗的承诺之一。在过去的一个世纪里，神经科学研究表明，在活动行为中，大脑的各个区域都需要处理感觉和运动信息。这些包括感觉和运动系统，也包括小脑和基底神经节等区域。虽然对特定回路的潜在过程已经有了部分了解，但理想情况下，全面了解需要对行为动物大脑中的神经活动进行记录。这种类型的实验在过去是不可能的，原因有两个：1。神经记录技术需要记录设备和神经组织之间的高度稳定性；因此，大多数实验都涉及严格限制或麻醉的动物，这阻止了有意义的大脑/环境相互作用。2. 典型的大脑记录要么局限于细胞分辨率下的少量神经元，要么局限于低空间和/或时间分辨率下的大面积脑组织的间接记录。为了解决这些挑战，将结合实验和计算神经科学的先进技术。首先，一个游动的斑马鱼幼虫的虚拟现实；这将允许我们从一个不运动的大脑中记录，但允许有效的行为。第二种是光片显微镜，这种技术可以同时对斑马鱼大脑中的10000个神经元进行成像。第三，分布式计算技术，这将使我们能够分析从这些实验中获得的大量数据集（每次试验高达1tb）。我们将使用这些工具来解决关于行为动物大脑功能的三个基本问题。首先，当动物积极地与世界接触时，大脑接收到两种类型的感觉输入：由外部世界变化引起的感觉输入，例如，当水流过鱼的视网膜时，鱼所经历的光流；以及由它们自己的行为引起的感觉输入，例如，鱼在游动时所经历的光流。这两种类型的输入传递不同类型的信息，但同时到达视网膜。因此，我们要问的一个核心问题是，让鱼区分它们的全脑回路是什么。其次，当动物自己的行为引起的感官输入不符合它们的期望时，它们会很容易地调整自己的行为。例如，当水粘度的变化导致它们游动的实际结果与预期结果不匹配时，也就是说，当它们的游动没有像预期的那样推动它们游动时，鱼类就会调节游动的强度。我们会问是什么分布式神经回路让鱼检测到这些不匹配的错误。我们将把我们的结果结合起来，在主动游动的鱼中产生一个生物学上合理的闭环控制模型，该模型再现了实验观察结果，并可用于启发机器人控制系统。该项目将开发新技术来记录和分析大型神经数据集，并为成功的积极行为所必需的大脑功能的分布式和动态性质提供独特的见解。

Most of what we do on a day-to-day basis involves the ongoing and fluid coordination between our senses and our actions. For example, making a cup of tea involves processing a constant stream of visual and tactile (touch) information to continuously correct how we move our muscles in order to avoid spilling milk or breaking a mug. Our brains coordinate this flow of sensory to motor information effortlessly, yet it is an ability that modern engineering still cannot rival; think of the clumsy robots in this year's DARPA Robotics Challenge. The ability to coordinate sensory input and motor actions is also often impaired in diseases like Parkinson's and dyspraxia. Understanding this ability is therefore not only a central goal of modern neuroscience but also one that promises to deliver advances for engineering and facilitate the treatment of disease.Over the last century, neuroscience research has revealed that areas across the brain are required to process sensory and motor information during active behaviours. These include sensory and motor systems but also areas such as the cerebellum and the basal ganglia. While a partial understanding of the underlying processes in specific circuits has been achieved, a full understanding would ideally require recordings of the neural activity from across the brain in a behaving animal. This type of experiment has been impossible in the past for two reasons: 1. Neural recording techniques require a great degree of stability between recording devices and neural tissue; thus most experiments involve heavily restrained, or anesthetized, animals, which prevents meaningful brain/environment interactions. 2. Typical brain recordings have been limited to either small numbers neurons at cellular resolutions or indirect recordings from large areas of brain tissue at low spatial and/or temporal resolution. To address these challenges will combine advanced techniques in experimental and computational neuroscience. First, a virtual reality for a swimming larval zebrafish; this will allow us to record from a non-moving brain but allow fictive behaviour. Second, light-sheet microscopy, a technique that can simultaneously image 10000's of neurons from across the zebrafish brain. Third, distributed computing techniques, which will enable us to analyse the enormous data sets (upto a terabyte per trial) acquired from these experiments.We will use these tools to address three fundamental questions about brain function in behaving animals. First, when animals actively engage the world the brain receives two types of sensory input: Sensory input caused by changes in the external world, e.g. the optic flow experienced by a fish as water sweeps past its retina, and sensory input that is a consequence of their own actions, e.g. the optic flow experienced by the fish that results from its own swimming. These two types of inputs convey different types of information but arrive together on the retina. Thus a central question we will ask is what are the brain-wide circuits that allow the fish to distinguish between them. Second, animals readily adapt their behaviour when the sensory inputs caused by their own actions do not meet their expectations. For example, fish modulate the strength of swimming when changes in water viscosity cause a mismatch between the actual and expected consequences of their swimming, i.e., when their swimming does not propel them as far as they expect. We will ask what the distributed neural circuits are that allow fish to detect these mismatch errors. We will combine our results to produce a biologically plausible model of closed-loop control in an actively swimming fish that reproduces experimental observations and could be used to inspire robotic control systems.This project will develop new techniques to record and analyse large neural datasets and provide unique insight into the distributed and dynamic nature of brain function necessary for successful active behaviour.