Discovering dynamic computations from large-scale neural activity recordings

项目摘要/摘要 神经活动如何在局部微电路内和跨大脑区域协调以驱动行为 神经科学中的一个中心悬而未决的问题。大规模并行神经记录技术的最新进展 在复杂的行为过程中，神经学正在以单神经元的粒度生成动态活动图 和单峰分辨率。为了揭示这些大规模数据集中的基本动态特征，新的 迫切需要原则性和可伸缩性的计算方法。为了满足这一需求，我们将取消- 开发一个广泛适用的非参数推理框架来发现动态计算 从大规模的神经活动记录中。我们的框架寻求数据的动态模型，但是 与现有技术不同，它不需要先验模型假设。现有技术通常 使用简单的自组织模型测试数据，这往往会遗漏或扭曲Defifi的动态特征。相反，我们的 非参数方法探索整个空间的所有可能的动态搜索模型 与数据一致，从而消除了先验的猜测工作，模棱两可的模型比较 和模型诱导的偏差。我们的目标是开发优化算法，以有效地搜索 所有动态模型的空间，在GPU上实现这些算法以实现最大计算量 速度，并得出信息理论界限，以量化我们的计算方法的可靠性。至 演示我们的新方法如何帮助科学fic发现，我们将使用它们来检查决策- 顶叶和运动前皮质的相关活动。而不同的决策理论模型 已经被提出，但仍然不知道决策计算是如何在水平上实现的 单个神经元和神经细胞群体。我们的分析将提供fi的第一个计算模型 决策直接植根于神经数据，协调了种群动态的稳定性和异质性。 单神经元反应的起源，揭示了不同皮质层决策计算的差异， 并确定兴奋性神经元和抑制性神经元在决策相关动力学方面的差异。

Project Summary/Abstract How neural activity is coordinated within local microcircuits and across brain regions to drive behavior is a central open question in neuroscience. Recent advances in massively-parallel neural recording tech- nologies are producing dynamic activity maps during complex behaviors, with single-neuron granularity and single-spike resolution. To reveal fundamental dynamic features in these large-scale datasets, new principled and scalable computational methods are urgently needed. To address this need, we will de- velop a broadly applicable, non-parametric inference framework for discovering dynamic computations from large-scale neural activity recordings. Our framework seeks a dynamical model of the data, but unlike existing techniques, does not require a priori model assumptions. Existing techniques commonly fit data with simple ad hoc models, which often miss or distort defining dynamic features. Instead, our non-parametric approach explores the entire space of all possible dynamics in search for the model consistent with the data, and thereby eliminates a priori guess work, ambiguous model comparisons and model-induced biases. We aim to develop optimization algorithms to effectively search through the space of all dynamical models, implement these algorithms on GPUs to achieve maximal computational speed, and derive information-theoretic bounds to quantify reliability of our computational methods. To demonstrate how our novel methods aid scientific discovery, we will employ them to examine decision- related activity in parietal and premotor cortices. While different theoretical models of decision-making have been proposed, it still remains unknown how decision computations are implemented on the level of individual neurons and neural populations. Our analyses will offer the first computational models of decision-making rooted directly in neural data, reconcile stability of population dynamics with hetero- geneity of single-neuron responses, reveal differences in decision-computations across cortical layers, and identify differences in decision-related dynamics of excitatory vs. inhibitory neurons.