CAREER: Building interpretable models of neural population activity through view-invariant representation learning and alignment

当我们用手碰一杯水，听树叶沙沙作响，或下一盘象棋时，大脑会发生什么？在我们的大多数经验中，感知和感觉是通过分布在大脑中的大规模神经元回路的活动来协调的。虽然神经记录的新进展扩大了我们测量大量（数百或数千）神经元活动的能力，但通过神经记录解析以“读出”意图或行为仍然是一个突出的挑战。本CAREER提案的目标是开发新的机器学习方法来学习神经活动和复杂行为之间的鲁棒映射。有了从大脑到行为的新方法，就有可能更好地理解神经计算，比较个体之间的神经活动，并创建捕捉大脑不断变化的本质的动态模型。该项目将分为三个目标，每个目标都侧重于开发方法，以解决构建大脑和行为之间映射的关键挑战。在目标1中，该项目将开发从神经群体活动中学习表征的新方法，重点是通过自监督和对比学习方法构建不变性。在目标2中，该项目将专注于跨多个神经记录共同学习表征的问题，并使用该技术来理解个体之间的共同因素和差异。在目标3中，该项目将开发提取动态潜在因素的方法，这些因素可以在更长的时间尺度上模拟表征的转变，并将其应用于健康衰老和神经退行性疾病的研究。该项目将开发机器学习框架和理论，用于从神经记录中学习稳健的表征，并提供新的方法来量化个体，随着时间，衰老或疾病的大脑活动变化。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

What happens in the brain when we move our hand to touch a glass of water, listen to the sound of rustling leaves, or play a game of chess? In most of our experiences, perception and sensation are orchestrated through the activity of large-scale circuits of neurons distributed throughout the brain. While new advances in neural recording have expanded our ability to measure the activity of large populations of (hundreds or thousands of) neurons, parsing through neural recordings to "read out" intent or behavior is still an outstanding challenge. The goal of this CAREER proposal is to develop new machine learning methods for learning robust mappings between neural activity and complex behavior. With new approaches that can go from the brain to behavior, it will be possible to better understand neural computation, compare neural activity between individuals, and create dynamic models that capture the ever-changing nature of the brain.The project will be organized into three aims, each of which focuses on development of methods to tackle key challenges in building a mapping between the brain and behavior. In Aim 1, the project will develop new methods for learning representations from neural population activity, with a focus on building invariances through self-supervised and contrastive learning methods. In Aim 2, the project will focus on the problem of learning representations jointly across multiple neural recordings and using this technology to understand common factors and differences across individuals. In Aim 3, the project will develop approaches to extract dynamic latent factors that model the shift in representations over longer time scales and apply them to study the study of healthy aging and neurodegenerative disease. This project will develop machine learning frameworks and theory for learning robust representations from neural recordings and provide new ways to quantify changes in brain activity across individuals, over time, aging, or disease.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.