From diverse dynamics to diverse computation via neural cell types

项目摘要 生物神经元网络的一个显著特征是其细胞类型的惊人多样性。少校， 国家协调的实验努力，包括大脑倡议的细胞普查网络(BICCN)和 艾伦研究所的细胞类型计划目前正在以新的和非常高的水平揭示这种细胞多样性 决心的力量。例如，仅在小鼠大脑皮质133的两个区域，就已经确定了细胞类型， 跨区域共享多种类型！在不同物种中也观察到了类似的细胞类别。这些类型显示 显著的差异不仅是基因表达和连通性，而且还有膜、尖峰和突触 动力学。 这与神经网络中的大多数计算和理论学习模型形成了鲜明对比 通常只使用一种或少量的细胞类型。这个项目的目标是生产新的 帮助缩小这一差距的计算和理论工具。这将使我们和更广泛的社区能够 检验细胞类型特定的、不同种类的细胞和突触动力学的功能作用的假设： 可以利用它们来生成复杂的网络动态，从而更快或更准确地学习 具有复杂动态的输入或目标本身的任务。这样的任务在自然界中比比皆是 环境。 验证这一假设需要新的高通量计算工具来训练神经网络 解决任务的生物现实动力学和连接性，了解多样性的新理论工具 蜂窝动力学有助于网络计算，并为大规模蜂窝数据驱动的新应用做出贡献 模特们。首先，在拨款支持的科学软件工程师的专业知识下，将构建、测试和 发布一个软件包，灵活地实现单细胞和短期的异质动力学 突触动力学。我们计划使用一个非常流行的、免费的和开源的软件框架来 机器学习(Pytorch)。接下来，我们将建立网络动态指标，以帮助机械地 解释细胞和突触的异质性在影响学习中起什么作用，什么作用不重要 性能。最后，我们将这些工具与先前的细胞类型特定计算模型集成在一起 小鼠初级视皮层，基于艾伦研究所的大规模数据库，以验证上述假设。 具体地说，我们将重新确定实验观察到的细胞和细胞中的异质性水平 突触动力学有助于视觉微电路执行视觉计算的能力。

Project Summary A prominent feature of biological neuronal networks is the astonishing diversity of their cell types. Major, nationally coordinated experimental efforts, including the BRAIN Initiative’s Cell Census Network (BICCN) and the Allen Institute Cell Types programs, are currently revealing this cellular diversity at new and very high levels of resolution. For example, just across two areas of mouse cortex 133 cell types have been characterized, with many types shared across areas! Similar cell classes have been observed across species. These types show marked differences not only gene expression and connectivity, but also membrane, spiking, and synaptic dynamics. This is in sharp contrast to most computational and theoretical models of learning in neural networks, which generally make use of only one or a small number of cell types. The goal of this project is to produce new computational and theoretical tools to help close this gap. This will enable us, and the broader community, to test a hypothesis for the functional role of cell-type specific, heterogeneous cellular and synaptic dynamics: that they can be harnessed to generate complex network dynamics which allows faster or more accurate learning of tasks which themselves have inputs or objectives which have complex dynamics. Such tasks abound in natural environments. Testing this hypothesis requires new high-throughput computational tools to train neural networks with biologically realistic dynamics and connectivity to solve tasks, new theoretical tools to understand how diverse cellular dynamics contribute to network computation, and new application to large-scale, cellular data-driven models. First, with the expertise of a grant-supported scientific software engineer, will build, test, and disseminate a software package that flexibly implements heterogeneous dynamics of single cells and short-term synaptic dynamics. We plan to use a very popular, freely available and open source software framework for machine learning (Pytorch). Next, we will establish metrics of network dynamics that help to mechanistically explain what does -- and does not -- matter about cellular and synaptic heterogeneity in impacting learning performance. Finally, we will integrate these tools with a prior cell-type specific computational model of the mouse primary visual cortex, based on large scale Allen Institute databases, to test the hypothesis stated above. Specifically, we will newly determine whether experimentally observed levels of heterogeneity in cellular and synaptic dynamics contribute to the ability of visual microcircuits to perform visual computation.