Revealing the organization and functional significance of neural timescales inauditory cortex

项目摘要 人们非常善于通过声音来理解世界：在嘈杂的环境中理解语言 餐厅，挑出家人的声音，或识别熟悉的旋律。尽管我们拿着这些 能力是理所当然的，它们反映了令人印象深刻的生物工程计算壮举 很难在机器系统中复制。我的研究计划的长期目标是开发计算机 以及实验方法来反向工程大脑如何对自然声音进行语音编码并利用 这些进展有助于理解和帮助听力障碍的治疗。的核心挑战之一 对自然声音进行编码是它们在许多不同的时间尺度上构造的，从毫秒到秒，以及 甚至几分钟。大脑是如何跨越这些不同的时间尺度来从声音中获得意义的？ 回答这个问题一直很有挑战性，因为没有通用的测量方法 大脑中的神经时标。因此，我们对神经时标是如何变化的知之甚少。 在听觉皮质中进行组织，以及这种组织如何实现对自然声音的编码。要克服这些障碍 限制，我们开发了一个简单的实验范式(“时间语境不变性”或TCI范式) 估计任何感觉反应的时间整合周期：刺激改变的时间窗口 回应。我们将TCI方法应用于人类皮层脑电和动物生理学。 记录，以揭示神经时标在区域和单细胞水平的组织(目标一)。引航员 我们的分析数据表明，时间尺度是按层次组织的，高阶区域显示 整合期大大延长。为了探索这个时间尺度层次结构的功能意义，我们 将TCI与非常适合描述自然声音的计算技术结合在一起(AIM II)。我们测试是否 增加的积分周期实现了更强的抗噪语音表示(Aim IIA)，无论区域 随着更长的积分周期编码自然声音的高阶属性(Aim IIB&IIC)，是否存在 语音或音乐等重要声音类别的专用集成期(AIM IID)，以及是否 皮层整合周期可以通过它们对特征作出反应的持续时间来解释(AIM IIE)。在 在进行这项研究的过程中，我将在两个关键方面进行培训：(1)ECoG，这是唯一的方法 用空间和时间的精确度来理解神经时标在人脑中是如何组织的 (2)深度神经网络(DNN)是唯一能够执行具有挑战性的感知任务的模型 并预测高阶大脑皮层区域的神经反应。在完成这次培训后，我将 有一套独特的实验(功能磁共振成像、脑电地形图、心理物理学)和计算技能(数据驱动统计 建模和假设驱动的DNN建模)，这将促进我向独立调查员的转变。

Project Summary People are remarkably adept at making sense of the world through sound: understanding speech in a noisy restaurant, picking out the voice of a family member, or recognizing a familiar melody. Although we take these abilities for granted, they reflect impressive computational feats of biological engineering that are remarkably difficult to replicate in machine systems. The long-term goal of my research program is to develop computational and experimental methods to reverse-engineer how the brain codes natural sounds like speech and to exploit these advances to understand and aid in the treatment of hearing impairment. One of the central challenges of coding natural sounds is that they are structured at many different timescales from milliseconds to seconds and even minutes. How does the brain integrate across these diverse timescales to derive meaning from sound? Answering this question has been challenging because there are no general-purpose methods for measuring neural timescales in the brain. As a consequence, we know relatively little about how neural timescales are organized in auditory cortex and how this organization enables the coding of natural sounds. To overcome these limitations, we develop a simple experimental paradigm (the “temporal context invariance” or TCI paradigm) for estimating the temporal integration period of any sensory response: the time window during which stimuli alter the response. We apply the TCI method to human electrocorticography (ECoG) and animal physiology recordings to reveal the organization of neural timescales at both the region and single-cell level (Aim I). Pilot data from our analyses reveal that timescales are organized hierarchically, with higher-order regions showing substantially longer integration periods. To explore the functional significance of this timescale hierarchy, we couple TCI with computational techniques well-suited for characterizing natural sounds (Aim II). We test whether increased integration periods enable a more noise-robust representation of speech (Aim IIA), whether regions with longer integration periods code higher-order properties of natural sounds (Aim IIB&IIC), whether there are dedicated integration periods for important sounds categories like speech or music (Aim IID), and whether cortical integration periods can be explained by the duration of the features they respond to (Aim IIE). In the process of conducting this research, I will be trained in two critical areas: (1) ECoG, which is the only method with the spatial and temporal precision to understand how neural timescales are organized in the human brain (2) deep neural networks (DNN) which are the only models able to perform challenging perceptual tasks at human levels and predict neural responses in higher-order cortical regions. After completing this training, I will have a unique set of experimental (fMRI, ECoG, psychophysics) and computational skills (data-driven statistical modeling and hypothesis-driven DNN modeling), which will facilitate my transition to an independent investigator.