RI: Small: Large-Scale Machine Learning for Connectomics

近年来，大规模图像数据分析已成为自然科学研究特别是神经科学领域的一个关键瓶颈。自动化数据采集的技术进步使收集tb和pb大小的数据集成为可能。手动提取这些数据集中包含的丰富信息需要大量的人力；从电子显微镜数据中重建完整果蝇大脑或小鼠皮质柱的神经连接，这些关键任务将需要人类使用目前最先进的手工和半自动方法进行一万年的劳动。改进的自动图像分析工具可能对神经科学界直接有用，可以从显微镜数据中大规模密集地重建神经回路，其中跟踪每个神经元过程的形态，识别细胞之间的所有化学突触连接，从而绘制神经组织中包含的电路的完整“接线图”。这种重建有可能从根本上影响对神经回路的理解，使大脑结构的竞争模型最终得到严格的实验验证或证伪。数据集的庞大规模，对高精度的需求以避免对数据得出错误的科学结论，以及对精心校准的置信度度量的需求，以限制必须花费人工验证算法输出的时间，这些都是现有分割方法无法很好地解决的重大挑战。研究人员建议(i)开发卷积位置敏感哈希的有效算法，这是一种新颖的位置敏感哈希技术的推广，适用于来自更大数据量的密集重叠补丁的高度适用设置。（ii）为稀疏编码的重叠补丁和卷积变体开发有效的算法，旨在扩展到非常大的数据集、过滤器大小和过滤器数量。提出的卷积位置敏感散列方法将用于实现这一点。（iii）开发利用(i)和（ii）分割电子显微镜数据的算法，并与现有的分割方法进行经验比较。所有提出的方法都具有高度可扩展性，可以在大型计算集群上执行，以处理大型训练和测试数据集。此外，由于所提出的方法允许数据的显式表示，因此它们有望比参数方法更好地校准，例如目前实现最佳精度的基于神经网络的电子显微镜数据分割方法。

Large-scale image data analysis has in recent years become a key bottleneck in natural science research, particularly in the field of neuroscience. Technological advances in automated data acquisition have enabled the collection of terabyte and petabyte-size datasets. Extracting the rich information contained in these datasets manually would require an inordinate amount of human labor; reconstructing the neural connectivity in a complete fruitfly brain or cortical column of a mouse from electron microscopy data, key tasks of interest, would require ten thousand years of human labor using current state-of-the-art manual and semi-automated approaches. Improved automated image analysis tools are likely to be directly useful to the neuroscience community, enabling large-scale dense reconstruction of neural circuits from microscopy data, in which the morphology of every neuronal process is traced and all chemical synaptic connections between cells are identified, thereby mapping the complete "wiring diagram" of the circuit contained in the neural tissue. Such reconstructions have the potential to fundamentally impact the understanding of neural circuits by enabling competing models of brain architecture to finally be rigorously verified or falsified experimentally.The large size of the datasets, the need for high accuracy to avoid incorrect scientific conclusions being drawn about the data, and the need for well-calibrated confidence measures in order to limit the time that must be spent manually verifying the output of algorithms, are all substantial challenges not well-addressed by existing segmentation methods. The investigators propose to (i) Develop efficient algorithms for convolutional locality-sensitive hashing, a novel generalization of locality-sensitive hashing techniques to the highly applicable setting of dense overlapping patches from a larger data volume. (ii) Develop efficient algorithms for the overlapping patch and convolutional variants of sparse coding designed to scale to very large datasets, filter sizes and numbers of filters. The proposed convolutional locality-sensitive hashing approach will be employed to enable this. (iii) Develop algorithms that leverage (i) and (ii) to segment electron microscopy data, and compare empirically to existing segmentation methods. All of the proposed methods are highly scalable to executions on large compute clusters in order to handle large training and test datasets. Furthermore, since the proposed methods allow explicit representation of the data, they are expected to be better calibrated than parametric methods such as the existing neural network-based methods for segmentation of electron microscopy data that currently achieve the best accuracy.