CRCNS: Resolving human face perception with novel MEG source localization methods

对一张脸的短暂一瞥，很快就会显示出关于我们面前这个人的丰富的多维信息。这一令人印象深刻的计算壮举是如何实现的？最近修订的人脸加工神经框架表明，对人脸形状信息的感知，即人脸不变特征，如性别、年龄和身份，是通过腹侧视觉通路处理的，包括枕面部区域、梭状面部区域和前颞叶面部区域。然而，来自功能磁共振成像的证据仍然不清楚何时、何地以及如何提取特定的面部维度，如年龄、性别和身份。复杂计算的一个关键属性是，它是通过阶段进行的，因此随着时间的推移而展开。我们最近在一项MEG研究中调查了面孔感知的计算阶段(Dobs等人，自然通讯，2019年)，发现性别和年龄在身份信息之前被提取。然而，由于目前脑磁源空间定位方法的局限性，这种时间信息还没有与fMRI提供的空间信息联系起来。在这里，我们建议克服这些限制，通过开发定位脑磁源的新方法，利用我们团队在脑磁图和机器学习方面的专业知识，提供人脸计算如何在大脑中的时间和空间展开的全貌。在目标1中，我们将开发一种新的分析性脑磁图定位方法，称为交替投影，迭代地将焦点源匹配到脑磁图数据。在目标2中，我们将开发一种新的基于几何深度学习的数据驱动的脑磁图定位方法，通过学习大脑皮层流形的非欧几里德空间中的统计关系来重建分布式大脑皮层地图。在目标3中，我们将首先确定哪种方法最适合于使用fMRI人脸定位器作为基本事实来模拟人类的脑磁图人脸反应。然后，我们将提取空间和时间上准确的人脸处理图，以表征沿腹侧视觉通路提取年龄、性别和身份信息所需的计算步骤。对人脸处理的神经基础进行精确的计算表征，将是人类视觉和社会知觉基础研究的里程碑式成就。深入了解人类面部感知是如何完成的，可能会进一步为如何改进执行类似任务的人工智能系统提供线索。此外，这里开发的方法可能会增加脑磁图数据的能力，以回答有关人脑中神经计算的时空轨迹的问题。

A brief glimpse at a face quickly reveals rich multi-dimensional information about the person in front of us. How is this impressive computational feat accomplished? A recently revised neural framework for face processing suggests perception of face form information, i.e. face invariant features such as gender, age, and identity, are processed through the ventral visual pathway, comprising the occipital face area, fusiform face area, and anterior temporal lobe face area. However, evidence from fMRI remains equivocal about when, where, and how specific face dimensions of age, gender, and identity, are extracted. A key property of a complex computation is that it proceeds via stages and hence unfolds over time. We recently investigated the computational stages of face perception in a MEG study (Dobs et al., Nature Comms, 2019) and found that gender and age are extracted before identity information. However, this temporal information has yet to be linked to the spatial information available from fMRI because of limitations in current methods for spatial localization of MEG sources. Here, we propose to overcome these limitations and provide the full picture of how face computations unfold over both time and space in the brain by developing novel methods for localizing MEG sources, leveraging our team’s expertise in MEG and machine learning. In Aim 1 we will develop a new analytical MEG localization method called Alternating Projections that iteratively fits focal sources to the MEG data. In Aim 2 we will develop a novel data-driven MEG localization method based on geometric deep learning that reconstructs distributed cortical maps by learning statistical relationships in the non-Euclidean space of the cortical manifold. In Aim 3, we will first identify which method is most suitable to model human MEG face responses using fMRI face localizers as ground truth. We will then extract spatially and temporally accurate face processing maps to characterize the computational steps entailed in extracting age, gender, and identity information along the ventral visual pathway. A computationally precise characterization of the neural basis of face processing would be a landmark achievement for basic research in vision and social perception in humans. Insights into how face perception is accomplished in humans may further yield clues for how to improve AI systems conducting similar tasks. Further, the methods developed here may increase the power of MEG data to answer questions about the spatiotemporal trajectory of neural computation in the human brain.