Structure and Function of a Cubic Millimeter of Cortex: Crowdsourcing for Proofreading and Discovery

到2020年底，IARPA微米计划将以自动重建所有 一立方毫米的小鼠视皮层中的神经元，以及神经元的突触连接和 钙成像对视频刺激的反应。我们相信，这个数据集可能会成为应用最广泛的 这是大脑皮层回路领域的一个重要资源，但要实现这一潜力还需要做更多的工作。我们建议创建 一个名为PYR(发音为“PEER”)的在线社区，它以微米的结果为起点。 我们提出通过自动分割来检测和纠正自动分割中的剩余错误 算法和人类校对协同工作。自动错误检测和纠错将利用 手工设计的启发式算法以及深度学习和深度度量学习。校对将是 众包给非科学家志愿者，这是Eyewire首创的概念，但经过更新后可以处理1000倍的 更大的数据集和更准确的人工智能。我们预计PYR将产生更多的皮质 在数量级上，它们之间的联系超过了以往任何一项研究。这一信息将伴随着 80,000个神经元的钙成像视觉反应。有了如此庞大和丰富的数据集，主要的 瓶颈将成为科学发现。我们建议通过组建一个社区来众包发现 与PYR和彼此合作的研究人员。探索基础设施将包括“科学应用编程接口” 它允许通过Python函数编程访问主要数据和派生数据， “物化”连接体的系统，使研究人员可以使用校对结果 基本上为零延迟，版本控制系统将确保分析可重现。我们将创建一个 使社区能够贡献自己类型的注释的框架，这些注释将包含在 物化。最后，我们将构建交互式发现工具，将Science API查询与 视觉化。我们设想了一个研究人员社区，他们将建立在上述发现基础设施的基础上， 创建和共享自己的分析工具。他们将研究关于大脑皮层结构的各种问题。 和功能，包括许多我们根本无法预见的东西。在未来，PYR可以合并其他皮质 数据集，与原始微米数据集相比，工程工作量要小得多。AS 有了微米，软件将在公共GitHub Repos中公开开发。该软件可以通过以下方式重复使用 其他人则要建立类似于PYR的系统。

At the end of 2020, the IARPA MICrONS program will conclude with an automated reconstruction of all neurons in a cubic millimeter of mouse visual cortex, along with the neurons’ synaptic connectivity and calcium-imaged responses to video stimuli. We believe that this dataset could become the most widely used resource in the field of cortical circuits, but more work is required to realize this potential. We propose to create an online community called Pyr (pronounced “peer”) that takes the result of MICrONS as its starting point. We propose to detect and correct the remaining errors in the automated segmentation through automated algorithms and human proofreading working in tandem. Automated error detection and correction will utilize hand-designed heuristics as well as deep learning and deep metric learning. Proofreading will be crowdsourced to nonscientist volunteers, a concept pioneered by Eyewire, but updated to handle a 1000x larger dataset and far more accurate artificial intelligence. We expect that Pyr will generate more cortical connections than any previous study, by orders of magnitude. This information will be accompanied by calcium-imaged visual responses of 80,000 neurons. With such an enormous and rich dataset, the primary bottleneck will become scientific discovery. We propose to crowdsource discovery by assembling a community of researchers who collaborate with Pyr and each other. Discovery infrastructure will include a “Science API” that allows programmatic access to the primary and derived data through Python functions, an automated system for “materializing” the connectome that makes the result of proofreading available to researchers with essentially zero delay, and a versioning system will insure that analyses are reproducible. We will create a framework enabling the community to contribute their own kinds of annotations, which will be included in materialization. Finally, we will build interactive discovery tools that couple Science API queries with visualization. We envision a community of researchers who will build upon the above discovery infrastructure, creating and sharing analysis tools of their own. They will investigate diverse questions about cortical structure and function, including many that we cannot foresee at all. In the future, Pyr could incorporate other cortical datasets as they become available, with much less engineering effort than the original MICrONS dataset. As with MICrONS, software will be developed in the open in public github repos. This software could be reused by others to build systems similar to Pyr.