Unified, Scalable, and Reproducible Neurostatistical Software

项目摘要 现代神经科学的许多进步都依赖于大神经的电生理记录 动物种群(如数百个细胞)或动物行为的高分辨率测量 (例如，从视频)。这些数据集解锁了一系列真正的转型 科学机会，因为它们使我们能够得出关于个人的可靠的统计推断 动物受试者在时间上被精确封装的时刻。然而，这些统计模型 在计算机软件中实现是复杂的和不平凡的。在过去的十年里，一个 最初，神经数据科学和统计学这个新生的子领域急剧增长，产生了 多种多样的建模方法和庞大而支离破碎的“一次性”格局 支持单一统计建模方法的软件包。这一探索 不同的统计方法一直是，并将继续是 推动这一领域的发展。尽管如此，将现有模型整合为一个 统一可靠的软件包早就该出现了。此外，这一努力必须解决 神经和行为数据的规模呈指数级增长，以及 对工作流进行建模。为了满足这些需求，该应用程序将开发新的软件 神经科学中一套经过时间检验的统计模型的实施。至 为了适应呈指数级增长的数据大小，该软件将构建在最近 用于大规模机器学习的创新基础设施，包括灵活的程序 指定要横扫的GPU加速计算和云计算框架 在多台机器上并行显示模型参数。最后，我们将制定程序，以 神经科学实验室与格式化的原始数据集共享可重复的分析工作流 根据Brain Initiative的标准，包括构建可共享URL的新框架， 在任何Web浏览器中运行的交互式数据可视化。总而言之，这些新的 软件工具将使实验室能够快速迭代，从而加速神经科学的发现 关于内部分析，并以透明和可重复的方式分享。

Project Summary Many advances in modern neuroscience rely on electrophysiological recordings of large neural populations (e.g. many hundreds of cells) or high-resolution measurements of animal behavior (e.g. from video). These datasets have unlocked a wide range of genuinely transformational scientific opportunities, as they enable us to draw reliable statistical inferences about individual animal subjects at precisely encapsulated moments in time. However, these statistical models are complex and non-trivial to implement in computer software. Over the past decade, an initially nascent sub-field of neural data science and statistics grew precipitously, producing a broad array of modeling approaches and a voluminous, fractured landscape of “one-off” software packages that support a single statistical modeling approach. This exploration of diverse statistical methodologies has been, and will continue to be, a crucial component to advancing the field. Nevertheless, a concerted effort to consolidate existing models into a unified and reliable software package is long overdue. Moreover, this effort must address the exponentially growing scale of neural and behavioral data, as well as the escalating intricacy of modeling workflows. To address these needs, this application will develop novel software implementations of a curated set of time-tested statistical models in neuroscience. To accommodate the exponentially growing data sizes, this software will be built on top of recently innovated infrastructure for large-scale machine learning, including flexible procedures for specifying GPU-accelerated computations and cloud computing frameworks to sweep across model parameters in parallel across many machines. Finally, we will develop procedures for neuroscience labs to share reproducible analysis workflows alongside raw datasets formatted by BRAIN Initiative standards, including a novel framework for building URL-shareable, interactive data visualizations that operate within any web browser. Altogether, these new software tools will accelerate neuroscience discoveries by enabling laboratories to rapidly iterate on in-house analyses and share them in a manner that is transparent and reproducible.