Computational neuroimaging of human auditory cortex

仅仅通过倾听，人类就可以推断出周围世界的大量事物：谁在说话，家里的窗户是开着还是关着，或者他们的孩子在隔壁房间的地板上掉了什么。这种从声音中获取信息的能力是人类智力的核心组成部分，它是由从耳朵延伸到大脑的神经元处理的许多阶段实现的。尽管我们对耳朵如何将声音转化为电信号并将其发送到大脑已经了解很多，但大脑调节我们声音识别能力的机制仍然知之甚少。这些知识上的差距限制了我们开发机器系统的能力，这些系统可以复制我们的听力技能（例如用于机器人），或者理解听力困难的基础，如阅读障碍、听觉处理障碍或与年龄相关的听力损失。为了深入了解使听觉识别成为可能的神经元过程，将使用功能磁共振成像（fMRI）来研究大脑对声音的处理，这是一种非侵入性测量大脑活动的技术。大脑中测量的反应将与最先进的声音识别计算机算法产生的数字反应进行比较。这项研究将有助于揭示人类听觉智能的原理，其长期目标是为听力障碍提供更有效的机器算法和治疗方法。这项研究还将深入了解计算机音频算法的内部工作原理，促进工程、工业和神经科学之间的互动。该项目将通过传播操纵声音和创建音频数据集的新工具来促进其他研究工作，并将在计算神经科学方面招募和培训妇女和代表性不足的少数民族。初级听觉皮层的结构和功能方面已经很好地建立起来，并且有各种各样的可能延伸到初级听觉皮层之外的途径的建议。然而，我们对听觉皮层中使声音识别成为可能的转换知之甚少，也很少有计算模型来解释这种转换是如何发生的。本研究的目标是进行功能磁共振成像实验，揭示听觉皮层中可能有助于听觉识别的表征转换，使用功能磁共振成像反应来测试现有的听觉计算模型，并开发可以解释人类能力和神经元反应的新模型。功能性MRI将用于表征皮层反应，因为它允许一次测量整个听觉皮层，使比较听觉皮层不同区域（包括远离皮层表面的区域）的反应成为可能，从而探测区域之间的表征转换。利用最近“深度学习”的成功，将开发新的听觉计算模型，并使用新的基于综合的模型评估方法来测试它们与大脑的相关性。研究结果将有助于揭示听觉皮层是如何调节声音识别的。

Just by listening, humans can infer a vast array of things about the world around them: who is talking, whether a window in their house is open or shut, or what their child dropped on the floor in the next room. This ability to derive information from sound is a core component of human intelligence, and is enabled by many stages of neuronal processing extending from the ear into the brain. Although much is known about how the ears convert sound to electrical signals that are sent to the brain, the mechanisms by which the brain mediates our sound recognition abilities remains poorly understood. These gaps in knowledge limit our ability to develop machine systems that can replicate our listening skills (e.g. for use in robots) or to understand the basis of listening difficulties, as in disorders such as dyslexia or auditory processing disorder, or in age-related hearing loss. To gain insight into the neuronal processes that enable auditory recognition, the brain's processing of sound will be studied using fMRI, a technique to non-invasively measure brain activity. The responses measured in the brain will be compared to the numerical responses produced by state-of-the-art computer algorithms for sound recognition. The research will help reveal the principles of human auditory intelligence, with the long-term goals of enabling more effective machine algorithms and treatments for listening disorders. The research will also provide insight into the inner workings of computer audio algorithms, stimulating interaction between engineering, industry, and neuroscience. The project will facilitate other research efforts via the dissemination of new tools for manipulating sound and the creation of audio data sets, and will recruit and train women and underrepresented minorities in computational neuroscience.Aspects of the structure and function of primary auditory cortex are well established, and there are a variety of proposals for pathways that might extend out of primary auditory cortex. However, we know little about the transformations within the auditory cortex that enable sound recognition, and there are few computational models of how such transformations might occur. The goal of the proposed research is to conduct fMRI experiments that reveal representational transformations within auditory cortex that might contribute to auditory recognition, to use fMRI responses to test existing models of auditory computation, and to develop new models that can account for human abilities and neuronal responses. Functional MRI will be used to characterize cortical responses because it allows measurements from the entire auditory cortex at once, making it possible to compare responses in different regions of the auditory cortex (including those far from the cortical surface), and thus to probe for representational transformations between regions. New models of auditory computation will be developed by leveraging the recent successes of "deep learning", and their relevance to the brain will be tested using new synthesis-based methods for model evaluation. The results will help reveal how the auditory cortex mediates robust sound recognition.