Neural mechanisms and behavioral consequences of non-Gaussian likelihoods in sensorimotor learning

神经科学的一个核心目标是了解神经系统如何实现学习。 然而，尽管在动物和人类身上进行了多年的研究， 学习的过程以及大脑对它的执行仍然是初级的。因此， 在学习过程中收集的大量行为和神经数据以及数学和 对运动可塑性规律的生物学理解。该提案将制定统一的 理解大脑如何学习复杂技能的数学理论。理论框架 将在软件中实现，并将适用于各种感觉运动数据并对其进行验证。 主要的实验验证系统将是鸣禽，它提供了一个生理上可访问的 模型系统来研究感觉运动学习。我们在鸣禽系统中的目标是了解 单个声学参数的感觉运动学习-基频（音高）-已知是 由鸣禽的大脑精确调节。我们的中心假设是，学习是作为一个贝叶斯实现的， 推理，并且从当前贝叶斯先验分布中随机采样运动命令 输出是由前脑中的神经元网络协调的。大量利用这两种物质 理论和实验结果，两个具体的目标将测试这一假设。第一个目标将介绍一个 一种创新的新型计算模型，其中大脑使用贝叶斯推理的迭代过程 根据感官反馈重塑行为。将使用人口平均值验证模型 动物行为第二个目标将分析从个体动物和单个神经元记录的数据， 行为动物来识别感觉运动学习的生物学机制。在整个过程中， 设计、测试和公开软件，使社区的其他成员能够应用我们的新工具 到他们自己的数据。我们的方法是创新的，因为它将提供一个统一的框架， 跨任务和物种的各种行为和神经研究的结果。这些研究 意义重大，因为更好地理解感觉运动学习的机制有助于 利用复杂行为的可塑性设计康复策略。

A central goal of neuroscience is to understand how learning is implemented by the nervous system. However, despite years of studies in animals and humans, our understanding of both the computational basis of learning and its implementation by the brain is still rudimentary. A critical gap therefore exists between the large amount of behavioral and neural data that has been collected during learning and a mathematical and biological understanding of the rules governing motor plasticity. This proposal will develop a unified mathematical theory for understanding how the brain learns complex skills. The theoretical framework will be implemented in software and will be applicable to and validated on a wide variety of sensorimotor data. The primary experimental validation system will be songbirds, which provide a physiologically accessible model system to investigate sensorimotor learning. Our objective in the songbird system is to understand sensorimotor learning of a single acoustic parameter – fundamental frequency (pitch) – which is known to be precisely regulated by the songbird brain. Our central hypothesis is that learning is implemented as a Bayesian inference, and that the stochastic sampling of motor commands from the current Bayesian a priori distribution of outputs is coordinated by a network of neurons in the forebrain. Drawing on a large quantity of both theoretical and experimental results, two specific aims will test this hypothesis. The first aim will introduce an innovative new class of computational model in which the brain uses an iterative process of Bayesian inference to reshape behavior in response to sensory feedback. The models will be validated using population-averaged animal behavior. The second aim will analyze data recorded from individual animals and single neurons in behaving animals to identify the biological mechanisms underlying sensorimotor learning. Throughout, we will design, test, and make public software that will allow other members of the community to apply our novel tools to their own data. Our approach is innovative because it will provide a unified framework for understanding the results of a wide variety of behavioral and neural studies across both tasks and species. These studies are significant because a better understanding of the mechanisms underlying sensorimotor learning could aid in the design of rehabilitative strategies that exploit the plasticity of complex behavior.