Cocktail Party Problem: Perspective on Neurobiology of Auditory Scene Analysis

A.目标和意义尽管在过去的几十年里计算技术取得了巨大的进步，但仍然有许多任务对孩子来说很容易，而对先进的计算机系统来说却很困难。大多数现有系统面临的一个特殊挑战是处理复杂的声学环境、背景噪音和相互竞争的说话者：这是鸡尾酒会上经常遇到的挑战（Cherry, 1953），正式称为听觉场景分析（Bregman, 1990）。该领域的进展具有巨大的影响和长期利益，涵盖医疗，工业，军事和机器人领域；以及改善沟通辅助设备（助听器，人工耳蜗，基于语音的人机界面），为感觉受损和老化的大脑。尽管它对工程和感知科学都很重要，但听觉场景分析的神经基础研究仍处于起步阶段。这一领域尤其受到缺乏整合理论的挑战，这些理论将我们对场景分析的感知基础的知识与听觉通路各个阶段的神经机制结合起来。由于问题的本质，神经回路的作用是复杂的，并且是设计成多尺度的。本研究的目的是为场景建模分析提供一个系统的视角，该视角整合了单神经元水平、种群水平和跨区域相互作用的机制。提出的理论的智力价值在于阐明了具体的机制和计算规则；促进其在工程系统中的集成，并能够生成新的可测试的预测。该提案调查了一个关键假设，即对复杂声音特征的关注实例化了与该特征一致的所有元素，从而将它们绑定在一起作为一个感知“对象”或流。这种“绑定假说”需要三个尺度的分析：微观层面的复杂声音映射成多维的皮层特征表示；皮层神经元群体相关活动的中观水平一致性分析以及调节听觉客体形成的注意和期望的宏观反馈过程。我们将在多尺度计算框架内制定这一假设，为听觉场景分析的神经基础提供统一的理论。本项目的三个核心研究目标是利用计算和生理学方法探索该模型的各个方面：目标1 .多尺度连贯模型：主要目标是将“绑定假设”作为听觉流的统一生物学理论，将多尺度感觉与认知皮质机制整合在一起。这一计算工作将纳入目标二和目标三的实验结果，生成可测试的预测，并提供有效的算法实现，以解决生物医学应用中的“鸡尾酒会问题”；目的二世。多尺度相干理论的生理研究：我们的目标是利用动物模型记录初级听觉和前额叶皮层的单单元（微观水平，中观水平）和跨区域（宏观水平）的生理活动，同时呈现足够复杂的声环境，以测试和完善计算模型；第三目标。用人类生理和感知测试改进相干理论：目标是使用脑磁图（MEG）和心理声学实验直接在人类受试者中测试模型的预测。我们将特别关注皮层机制在正常和衰老大脑场景分析中的作用。拟议的研究借鉴了一个跨学科团队的专业知识，整合了神经生物学和工程学。它的独特之处在于，它是第一次假设连贯性在场景分析问题中的作用，并研究听觉流实验中整合皮层和注意机制的“绑定假设”。此外，通过直接在人类受试者身上测试该理论，并比较正常和衰老的大脑（已知在鸡尾酒会环境中面临感知困难），我们希望更好地了解正常和故障状态下场景分析的神经基础，从而增强模型的翻译潜力。这项工作的更广泛的影响是提供了听觉流分离的通用和易于处理的模型，显著地促进了这种能力在工程系统中的集成。

DESCRIPTION (provided by applicant): A Multi-scale Perspective on the Neurobiology of Auditory Scene Analysis A. Aims and Significance Despite the enormous advances in computing technology over the last decades, there are stills many tasks that are easy for a child, yet difficult for advanced computer systems. A particular challenge to most existing systems is dealing with complex acoustic environments, background noises and competing talkers: A challenge often experienced in cocktail parties (Cherry, 1953) and formally referred to as auditory scene analysis (Bregman, 1990). Progress in this field has tremendous implications and long- term benefits covering the medical, industrial, military and robotics domains; as well as improving communication aids (hearing aids, cochlear implants, speech-based human-computer interfaces) for the sensory-impaired and aging brains. Despite its importance for both engineering and perceptual sciences, the study of the neural underpinnings of auditory scene analysis remains in its infancy. This field is particularly challenged by the lack of integrative theories which incorporate our knowledge of the perceptual bases of scene analysis with the neural mechanisms along various stages of the auditory pathway. Because of the nature of the problem, the neural circuitry at play is intricate and multi-scale by design. The objective of the proposed research is to provide a systems view to modeling scene analysis which integrates mechanisms at the single neuron level, population level and across area interactions. The intellectual merit of the proposed theory is to elucidate the specific mechanisms and computational rules at play; facilitate its integration in engineering systems and enable generating novel testable predictions. The proposal investigates the key hypothesis that attention to a feature of a complex sound instantiates all elements that are coherent with this feature, thus binding them together as one perceptual "object" or stream. This "binding hypothesis" requires three scales of analyses: a micro-level mapping of complex sounds into a multidimensional cortical feature representation; a meso-level coherence analysis correlating activity in populations of cortical neurons; and macro-level feedback processes of attention and expectations that mediate auditory object formation. We shall formulate this hypothesis within a multi-scale computational framework that provides a unified theory for the neural underpinnings of auditory scene analysis. The three core research aims of this project explore all facets of this model employing computational and physiological approaches: Aim I. A multi-scale coherence model: The main goal is to formulate the "binding hypothesis" as a unified biologically plausible theory of auditory streaming, integrating multi-scale sensory with cognitive cortical mechanisms. This computational effort will incorporate findings from experiments in Aims II and III, generate testable predictions, as well as provide effective algorithmic implementations to tackle the "cocktail party problem" in biomedical applications; Aim II. Physiological investigations of the multi-scale coherence theory: Our aim is to use an animal model to record single-unit (micro-level, meso-level) and across area (macro-level) physiological activity in both primary auditory and prefrontal cortex, while presenting sufficiently complex acoustic environments so as to test and refine the computational model; Aim III. Refinement of the coherence theory with physiological and perceptual testing in humans: The objective is to directly test predictions from the model in human subjects, using magnetoencephalography (MEG) and psychoacoustic experiments. We shall particularly focus on the role of cortical mechanisms in scene analysis in normal and aging brains. The proposed research draws upon the expertise of a cross-disciplinary team integrating neurobiology and engineering. It is unique in that it is the first effort to postulate a role for coherence in the scene analysis problem, and to investigate the "binding hypothesis" integrating cortical and attention mechanisms in auditory streaming experiments. In addition, by testing the theory directly on human subjects and comparing normal and aging brains (known to face perceptual difficulties in cocktail party settings), we hope to better understand the neural underpinnings of scene analysis under their normal and malfunctioning states, hence enhancing the translational potential of the model. The broader impact of this effort is to provide versatile and tractable models of auditory stream segregation, significantly facilitating the integration of such capabilities in engineering systems.