Observable signatures of learning in neural circuits

学习是神经系统的一项重要功能，它源于神经元之间突触连接的可塑性。因此，神经科学的一项主要工作旨在了解突触可塑性的基本原理：确定可以预测突触强度何时变化以及变化程度的数学规则。虽然在脑组织切片中进行的实验可以测量突触的强度，以及它们在驱动可塑性的过程中的变化，但目前还没有可靠的方法在活体动物的大脑中进行同样的实验。因此，在我们最关心的条件下，神经科学家对通过突触可塑性学习是如何工作的缺乏理解。为了克服这个困难，我建议开发一个数学框架，揭示神经回路中不同突触可塑性规则的特征，这些特征可以使用通常从活体动物的大脑中收集的数据类型来观察。接下来，我将把这个数学框架应用到我的合作者在清醒行为动物的大脑中收集的数据中，以确定与神经数据最一致（或最不一致）的可塑性机制。

Learning is a critical function of the nervous system, and it arises from the plasticity of the synaptic connections between neurons. Accordingly, a major effort in neuroscience aims to understand the principles underlying synaptic plasticity: to identify mathematical rules that can predict when, and by how much, synaptic strengths will change. While experiments in slices of brain tissue can measure the strengths of synapses, and their changes during protocols that drive plasticity, there is currently no reliable way to do these same experiments in the brains of living animals. Thus, neuroscientists lack an understanding of how learning via synaptic plasticity works under the conditions that we care most about. To overcome this difficulty, I propose to develop a mathematical framework that will reveal signatures of different synaptic plasticity rules within a neural circuit, that can be observed using the types of data typically collected from the brains of living animals. Next, I will apply this mathematical framework to data collected by my collaborators in the brains of awake behaving animals, to identify the plasticity mechanisms that are most (or least) consistent with the neural data. For each possible synaptic plasticity mechanism, we will use analytical calculations, and simulations of neural circuits, to identify the mean-variance relationship, and the relationship between signal and noise correlations, in the neurons' activities. These quantities are often measured in the brains of awake behaving animals, using standard experimental methods (Utah arrays, Ca2+ imaging, etc.), both by my collaborators (e.g., Brain Observatory team at the Allen Institute for Brain Science), and by others who make their data publicly available (e.g., via the CRCNS repository). Next, we will develop a data-science method that will "demix" neural data into components with different mean-variance relationships and relationships between signal and noise correlations; this method will estimate the relative contributions of each of these components. Finally, we will apply this method to neural data. By identifying each component with the plasticity mechanism that yields the same mean-variance relationship and relationship between signal and noise correlations, we will estimate the relative contribution of each plasticity mechanism to the neural variability within each dataset. By identifying the mechanisms underlying learning, this work may lead to better treatments for those with learning disorders. Moreover, this work could lead to subsequent advances in machine learning (ML): by implementing these mechanisms in next-generation ML systems, developers can create more biorealistic ML. Inasmuch as today's ML algorithms are inspired by our current understanding of neural information processing, and are already impressively powerful, we anticipate that these next-generation ML systems could have substantial impacts on the broader community.