Nonparametric Bayes Methods for Big Data in Neuroscience

描述（由申请人提供）：我正在申请通过BD 2K计划的指导职业发展，以获得过渡到独立研究职业所需的技能和专业知识，开发系统和认知神经科学中“大数据”分析方法。在我的博士学位之后。在接受理论物理学培训后，我转向了计算神经科学，在那里我专注于奖励和决策的神经生理学问题，特别是强化学习和选择行为的模型。在过去的五年里，我也获得了在人类手术患者和非人类灵长类动物的电生理记录方面的丰富经验，加深了我对分析真实的神经科学数据所涉及的困难的理解。在这段时间里，我已经确信，在未来十年中，神经科学面临的最紧迫的挑战将是我们如何处理，分析和综合新技术提供的快速增长的数据量的问题，当我过渡到教师水平时，我正在寻求将我自己的研究计划导向这些目标。要做到这一点，我需要补充我强大的定量背景和电生理记录技能，并在机器学习，信号处理和功能性磁共振成像（fMRI）数据分析方面进行专门培训。我专注于第一个，因为数据分析的统计数据是必不可少的 这是因为：第一，了解我们处理和获取数据的方法与我们如何分析它们一样重要;第二，因为fMRI数据不仅是最容易获得的大型数据集之一，而且对fMRI数据的有效分析将立即具有临床应用。为了这个项目，我组建了一个导师团队，他们在这三个领域拥有强大而重叠的专业知识。这些导师致力于支持我转向专注于大数据研究，这种方法建立在我与杜克实验室的多个现有合作基础上。我的最终目标是领导一个实验室，在这个实验室里，我将在获奖期间获得的技能和培训应用于开发计算方法，这些方法将利用大数据的力量来回答认知和转化神经科学中的基本问题。环境杜克大学在神经科学和大数据研究方面都拥有杰出的资源。其跨学科的大数据努力，信息倡议在杜克，汇集了来自统计，计算机科学和电气工程与遗传学，神经科学和社会科学的研究人员，以促进跨学科的合作。我所在的杜克脑科学研究所由杜克脑科学领域的150多名教师组成，从临床医生到生物医学工程师。我将由大卫邓森博士指导，他是机器学习贝叶斯统计方法的公认领导者，沿着还有劳伦斯·卡林博士和吉列尔莫·萨皮罗博士，他们是信号和图像处理以及机器学习方面的专家，也是邓森博士的经常合作者。此外，功能磁共振成像专家、领先神经成像教科书的作者斯科特·休特尔博士将监督我在功能磁共振成像数据分析方面的培训。此外，我还将获得来自杜克大学大量不同实验室的数据，包括美国最大的神经成像数据集之一。最重要的是，杜克大学将全力支持我，提供必要的资源和时间，让我能够接受职业发展奖中列出的培训。Research.每年，每四个成年人中就有一个患有可诊断的精神障碍，每25个成年人中就有一个患有严重的精神疾病。然而，我们预测精神疾病发作的能力--甚至我们理解其在大脑中的影响的能力--一直受到这样一种认识的限制，即这些疾病主要不是独立单位的紊乱，而是病理性大脑激活的模式。然而，我们目前缺乏对神经网络内活动模式的有意义的表征，因此缺乏有效讨论、发现和处理它们的能力。然而，我们表征和检测这些模式的能力的提高将产生重大的临床影响。因此，在我的指导团队的指导下，我建议使用机器学习的方法来描述神经科学数据集中的网络活动模式。因为许多精神疾病的典型特征要么是患者与世界刺激之间的病理关系（后 创伤应激障碍、饮食失调）或思维障碍的内在模式（重度抑郁症、强迫症），我专注于模式检测的三个关键问题：1）大脑如何编码复杂的、非结构化的刺激？2)健康和疾病的内在大脑活动模式的基本组成部分是什么？3)如何将 大脑活动的变化对行为状态变化的反应？我的方法利用了贝叶斯非参数方法的最新进展，以及快速变分推理方法，可以很好地扩展到大型数据集。此外，因为我将使用的数据集，fMRI和电生理数据，是更大类别的多通道时间序列数据的特殊例子，结果将更广泛地应用于神经科学和其他领域的其他类型的数据。

DESCRIPTION (provided by applicant): I am applying for mentored career development through the BD2K initiative to gain the skills and expertise necessary to transition to an independent research career developing methods for the analysis of "big data" in systems and cognitive neuroscience. Following my Ph.D. training in theoretical physics, I transitioned into computational neuroscience, where I have focused on problems in the neurophysiology of reward and decision-making, particularly models of reinforcement learning and choice behavior. For the last five years, I have also gained extensive experience in electrophysiological recording in both human surgical patients and non-human primates, deepening my appreciation of the difficulties involved in analyzing real neuroscience data. During this time, I have become convinced that the single most pressing challenge for neuroscience in the next decade will be the problem of how we process, analyze, and synthesize the rapidly expanding volumes of data made available by new technologies, and as I transition to the faculty level, I am seeking to orient my own research program toward these goals. To do so, I will need to complement my strong quantitative background and electrophysiological recording skills with specific training in machine learning, signal processing, and analysis of data from functional magnetic resonance imaging (fMRI). I am focusing on the first because the statistics of data analysis are an essential core competency for any big data researcher; on the second because understanding the methods by which we process and acquire data are as essential as how we analyze them; and on the third because not only are fMRI data among the most readily available large datasets, but effective analysis of fMRI data will have immediate clinical applications. For this project, I have assembled a team of mentors with strong and overlapping expertise in these three areas. These mentors have committed to support my transition to a focus on big data research, an approach that builds on multiple existing collaborations I have with laboratories at Duke. My ultimate goal is to head a lab in which I apply the skills and training I acquire during the award period to developing computational methods that will harness the power of big data to answer fundamental questions in cognitive and translational neuroscience. Environment. Duke University is home to outstanding resources in both neuroscience and big data research. Its interdisciplinary big data effort, the Information Initiative at Duke, brings together researchers from statistics, computer science, and electrical engineering with those in genetics, neuroscience, and social science to facilitate collaboration across the disciplines. The Duke Institute for Brain Sciences, with which I am affiliated, comprises over 150 faculty across the brain sciences at Duke, from clinicians to biomedical engineers. I will be mentored by Dr. David Dunson, a recognized leader in Bayesian statistical methods for machine learning, along with Dr. Lawrence Carin and Dr. Guillermo Sapiro, experts in signal and image processing and machine learning and frequent collaborators with Dr. Dunson. In addition Dr. Scott Huettel, an expert in fMRI and author of a leading neuroimaging textbook, will oversee my training in fMRI data analysis. Moreover, I will have access to data from a large and diverse pool of laboratories at Duke, including one of the largest neuroimaging datasets in the country. Most importantly, Duke is fully committed to supporting me with the resources and time necessary to pursue the training outlined in this career development award. Research. Each year, one in four adults suffers from a diagnosable mental disorder, with 1 in 25 suffering from a serious mental illness. Yet our ability to anticipate the onset of mental illness - even our ability to understand its effets within the brain - has been limited by the recognition that these diseases are not primarily disorders of independent units, but patterns of pathological brain activation. However, we currently lack a meaningful characterization of patterns of activity within neural networks, and thus the ability to discuss, discover, and treat them effectively. Yet an improvement in our abilit to characterize and detect these patterns would result in major clinical impact. Therefore, under the guidance of my mentoring team, I propose to characterize patterns of network activity in neuroscience datasets using methods from machine learning. Because many mental illnesses are typified either by a pathological relationship between sufferers and stimuli in the world (post traumatic stress disorder, eating disorders) or intrinsic patterns of disordered thought (major depression, obsessive-compulsive disorder), I focus on three key questions for pattern detection: 1) How does the brain encode complex, unstructured stimuli? 2) What are the basic building blocks of healthy and diseased patterns of intrinsic brain activity? 3) How do patterns of brain activity change in response to changes in behavioral state? My approach makes use of recent advances in Bayesian nonparametric methods, as well as fast variational inference approaches that scale well to large datasets. In addition, because the datasets I will use, fMRI and electrophysiology data, are particular examples of the much larger class of multichannel time series data, the results will apply more broadly to other types of data, in neuroscience and beyond.