Machine Learning for Analyzing State Dependent Neuronal Network Dynamics

钙成像允许在单个宽视野中记录100个神经元， 到极高维度的数据。目前的分析标准采用描述性统计 将神经元反应总结为单一的定量指标， 个体细胞和局部网络的动力学。相比之下，机器学习，尤其是 降维模型，提供了更细致的分析，考虑了时间 细胞之间的模式和分组。虽然以前的工作试图减少神经元 活动非常低维流形，这些方法的结果是很难输出 明白在这项工作中，我们采用非负矩阵分解（NMF），一个容易的 一种分析神经网络动力学位移的可解释降维方法 不同的实验背景下产生的不同的结果。我们将应用我们的框架来研究 两种不同背景下的神经元网络动力学：1）初级躯体感觉皮层 (S1)在增加浓度的麻醉下，和2）光遗传学期间的海马 刺激神经元的记忆编码集合。我们成功地适应了， 描述了一系列降维方法，并证明了NMF是一种 从钙记录中提取底层结构的上级方法。初步分析表明， 提取有序的、低维的、用传统统计学无法检测的内部结构。 这项研究将在波士顿大学进行，利用众多的 多学科研究中心（系统神经科学中心，神经光子学中心， 拉菲克B。哈里里计算和计算科学与工程研究所）。这些 学院，由高度多样化和知名的教师群体组成，创造了一个高度 跨学科研究的协作环境，使科学家能够追求有趣的 不直接涉及他们专业知识的问题。此外，结合学术培训， 技术技能的发展，分析问题的解决，科学交流， 专业发展和一贯的指导将确保项目具有最高的 成功的可能性。

Calcium imaging allows recording from 100s of neurons in a single wide field of view, giving rise to extremely high dimensionality data. Current analysis standards employ descriptive statistics that summarize neuronal responses into single quantitative metrics, discounting the temporal dynamics of individual cells and local networks. In contrast, machine learning, especially dimensionality reduction models, provide more nuanced analysis that considers the temporal patterns and groupings among cells. While previous work has attempted to reduce the neuronal activity to very low dimensional manifolds, these methods result in outputs that are difficult to understand. In this work, we adapt Non-Negative Matrix Factorization (NMF), an easily interpretable dimensionality reduction method to analyze shifts in neuronal network dynamics that arise as a function of different experimental contexts. We will apply our framework to study the neuronal network dynamics of two different contexts: 1) the primary somatosensory cortex (S1) under increasing concentrations of anesthesia, and 2) the hippocampus during optogenetic stimulation of memory-encoding ensembles of neurons. We have successfully adapted and characterized a series of dimensionality reduction methods and have demonstrated NMF is a superior method to extract underlying structure from calcium recordings. Initial analyses have extracted ordered, low-dimensional, internal structure not detectable with traditional statistics. This research will be conducted at Boston University, taking advantage of the numerous multidisciplinary research centers (Center for Systems Neuroscience, Neurophotonics Center, Rafik B. Hariri Institute for Computing and Computational Science & Engineering). These institutes, consisting of highly diverse and renowned groups of faculty, create a highly collaborative environment for interdisciplinary research, allowing scientists to pursue interesting questions not directly in their expertise. Further, a combination of academic training, development of technical skills, analytical problem solving, scientific communication, professional development, and consistent mentorship will ensure the project has the highest potential to succeed possible.