Dynamic network computations for foraging in an uncertain environment

 描述（由申请人提供）：大脑进化出复杂的循环网络，以在动态和不确定的世界中实现灵活的行为，但其计算策略和潜在机制仍然知之甚少。我们建议揭示觅食神经计算的网络基础，这是一项与行为学相关的行为任务，涉及感觉整合，空间导航，记忆和复杂的决策。我们将使用大规模的电子记录从六个相关的相互关联的区域（视觉皮层V4区，区域7A，内嗅皮层，海马，海马旁回，和前额皮质）的自由行为的猕猴。为了跟踪这些行为学相关的自然任务中使用的神经网络计算，我们将利用统计数据分析和神经计算理论的最新进展。首先，为了表征行为，我们将使用图形建模来建模与任务相关的感觉、运动和内部变量之间的关系。动物行为将在部分可观察马尔可夫决策过程（POMDP）的框架中建模，这些模型将提供关于动物使用哪些变量以及它们如何相互作用的预测。其次，一旦我们对行为相关变量进行建模，我们将使用现代数据分析技术从神经元反应模式中识别这些变量，从高维群体活动中提取低维任务相关信号。这些低维神经表征的时间序列将用于分析不同大脑区域之间的信号转换和流动，使用定向信息等措施。最后，我们将比较这些神经分析的预测觅食任务的规范模型。我们假设感觉和内部变量的神经表征将表现出与行为模型中相同的因果关系和时间关系。通过首次结合规范建模、神经群体信号的选择性降维和定向信息流的量化，我们将能够识别在复杂自然任务上执行神经计算的关键大脑区域内部和之间的转换。该团队项目旨在产生分布式神经群体编码的变革性观点，统一跨多个神经系统的行为学关键计算。

 DESCRIPTION (provided by applicant): The brain evolved complex recurrent networks to enable flexible behavior in a dynamic and uncertain world, but its computational strategies and underlying mechanisms remain poorly understood. We propose to uncover the network basis of neural computations in foraging, an ethologically relevant behavioral task that involves sensory integration, spatial navigation, memory, and complex decision-making. We will use large-scale electrical recordings from six relevant interconnected areas (visual cortical area V4, Area 7A, Entorhinal Cortex, Hippocampus, Parahippocampal gyrus, and Prefrontal Cortex) of freely behaving macaques. To track the neural network computations used in these ethologically relevant, natural tasks, we will exploit recent advances in both statistical data analysis and theories of neural computation. First, to characterize behavior, we will model relationships between task-relevant sensory, motor, and internal variables using graphical modeling. Animal behavior will be modeled in the framework of Partially Observable Markov Decision Processes (POMDP) and these models will provide predictions about which variables the animals use and how they interact. Second, once we have modeled the behaviorally relevant variables, we will use modern data analysis techniques to identify these variables from the patterns of neuronal responses, extracting the low- dimensional, task-relevant signals from the high-dimensional population activity. The time series of these low- dimensional neural representations will be used to analyze the transformation and flow of signals between different brain areas, using such measures as Directed Information. Finally, we will compare these neural analyses to predictions from the normative models of the foraging task. We hypothesize that neural representations of sensory and internal variables will exhibit the same causal and temporal relationships manifested in the behavioral model. By combining - for the first time - normative modeling, selective dimensionality reduction of neural population signals, and quantification of directed information flow, we will be able to identify the transformations within and between key brain areas that enact neural computations on complex natural tasks. The team project aims to produce a transformative view of distributed neural population coding, unifying ethologically crucial computations across multiple neural systems.