Behavioral Analysis and Modeling Core

摘要/摘要，核心D：行为分析和建模 这项提案的首要目标是了解内部国家如何影响决策，并确定 潜在的神经机制。行为分析和建模核心的开发、测试和 应用统计工具严格描述行为状态是实现这一目标的关键。这 协作将研究在三个不同时间尺度上定义的行为状态变化： 自发地投入和脱离一项任务，这些都是由于预期的变化而产生的 在任务期间，以及在几天内和跨天学习所产生的结果。这个核心的目标是开发 并扩展了用于从行为和视频数据中提取状态信息的新型开源分析工具 在这三个时间尺度上。首先，调查人员将确定支配选择行为的潜在状态，这 在实验过程中，使用基于隐马尔可夫模型的工具和泛化的 线性模型输出。此外，他们还将开发一种分层扩展的模型，以进行统计 利用拟议的实验项目产生的庞大的行为数据集。接下来，他们将推断 使用尖端视频分析方法在一次试验中改变的行为状态。特别是，他们将 应用最先进的无标记跟踪方法提取动物特征(爪子、舌头、鼻子、 等)从行为视频中，扩展这些方法以获得三维动物姿势的估计 (融合多个摄影机视图)。然后，他们会将无标记跟踪输出与非线性相结合 自动编码器压缩方法，以获得信息量更大的半监督、低维数据 视频数据的表示。使用机器学习方法，他们将对结果进行时间分割 表示以在每个试验中获得可解释的行为状态(例如，“休息”、“梳理”、“达到”)，适合 用于进一步的下游分析。最后，他们将开发新的工具来跟踪 学习的过程。决策战略在培训期间发展，包括在课程内和课程之间，以及 即使在训练有素的动物身上也会继续变化。为了描述这些状态的变化，调查人员将开发出 并应用将状态空间建模和强化学习方法相结合的新型统计模型来 分析在单个动物身上观察到的学习曲线，因为它们被训练来执行国际大脑 实验室决策任务。由此产生的框架将量化有多大的显著差异 在动物之间的学习曲线可以归因于识别的学习规则的差异，并将有助于识别 全脑录音中推断学习动力的神经关联。所有开发的软件工具都将 完全开源，并将通过公共、并行的云实施共享，以实现最大的可扩展性 和再现性。

Summary/Abstract, Core D: Behavioral Analysis and Modeling This proposal’s overarching goal is to understand how internal states influence decisions and to identify the underlying neural mechanisms. The Behavioral Analysis and Modeling Core’s development, testing, and application of statistical tools to rigorously characterize behavioral states is critical to achieving this goal. This collaboration will study behavioral state changes defined on three different time scales: those arising spontaneously with engagement and disengagement in a task, those resulting from changing expectations during the task, and those resulting from learning within and across days. The goals of this core are to develop and extend novel open-source analytical tools for extracting state information from behavioral and video data over these three timescales. First, the investigators will identify latent states governing choice behavior, which vary across trials within an experimental session, using tools based on a hidden Markov model with generalized linear model outputs. In addition, they will develop a hierarchical extension of the model to take statistical advantage of the vast behavioral dataset produced by the proposed experimental projects. Next, they will infer behavioral states that vary within a single trial using cutting-edge video analysis methods. In particular, they will apply state-of-the-art markerless tracking methods to extract the position of animal features (paws, tongue, nose, etc.) from behavioral video, and extend these methods to obtain estimates of animal pose in three dimensions (fusing multiple camera views). They will then combine the markerless tracking output with nonlinear autoencoder compression methods to obtain a more informative semi-supervised, low-dimensional data representation of the video data. Using machine learning methods, they will temporally segment the resulting representation to obtain interpretable behavioral states within each trial (e.g., “rest,” “groom,” “reach”), suitable for further downstream analyses. Finally, they will develop new tools to track the dynamics of behavior over the course of learning. Decision-making strategies evolve during training, both within and across sessions, and continue to vary even in well-trained animals. To characterize these state changes, the investigators will develop and apply novel statistical models that combine state-space modeling and reinforcement learning approaches to analyze the learning curves observed in individual animals as they are trained to perform the International Brain Laboratory decision-making task. The resulting framework will quantify how much of the pronounced differences in learning curves across animals can be attributed to differences in identified learning rules, and will help identify neural correlates of inferred learning dynamics in brainwide recordings. All software tools that are developed will be fully open source and will be shared via a public, parallelized cloud implementation for maximal scalability and reproducibility.