Neural Representations of Objects Across the Human Visual Pathway

人们擅长通过位置、大小、视点、光线和一般形式的变化来识别对象，而计算机识别系统在面对这种可变的、不可预测的情况时表现不佳。人类大脑究竟是如何解决物体识别的计算挑战的，目前还不是很清楚。要识别对象，必须将对象的特征、轮廓和部分集成到一个有组织的整体中，然后将这些对象形状的表示与存储在内存中的项进行匹配。不知何故，大脑可以通过提取物体的稳定、不变的属性来解决这个计算问题，而不考虑视网膜图像的表面变化。在美国国家科学基金会的支持下，范德比尔特大学的Frank Tang博士和他的同事们将使用功能磁共振成像(FMRI)和适应于机器学习的新模式分类方法来研究对象识别的神经基础。这些研究将确定人类视觉通路上的皮质活动模式代表了关于物体的哪种类型的信息，从对基本特征反应最好的低水平视觉区域到对复杂物体反应最好的高水平区域。这个项目不是只关注高级别的对象选择区域，而是强调一种不同的方法来理解对象的不变表示是如何形成的。研究将在视觉通路的每个阶段，从初级视觉皮质到前颞下区域，表征物体的神经表征，以确定物体表征如何从一个处理阶段转换到下一个处理阶段。这项研究将有助于揭示大脑是如何解决物体识别问题的，方法是通过跨越视觉通路的许多连续水平的过程，将原始的视网膜输入转换为对物体越来越灵活的表示。这些研究的结果将为当前的物体识别理论提供参考。了解这些神经基础对于理解在学习障碍、发育障碍或脑损伤(例如，发育性或获得性阅读障碍)的情况下，物体识别可能会出现什么问题是必要的。用于识别对象的计算机算法也可能会得到改进。

People excel at recognizing objects across changes in position, size, viewpoint, lighting and general form, whereas computer recognition systems perform poorly when faced with such variable, unpredictable situations. Exactly how the human brain solves the computational challenges of object recognition is not well understood. To recognize an object, one must integrate the features, contours, and parts of an object into an organized whole, and then match these representations of an object's shape to items stored in memory. Somehow, the brain can solve this computational problem by extracting the stable, invariant properties of objects while disregarding superficial variations in the retinal image. With support from the National Science Foundation, Dr. Frank Tong and his colleagues at Vanderbilt University will investigate the neural bases of object recognition using functional magnetic resonance imaging (fMRI) and novel pattern classification methods adapted from machine learning. These studies will determine what types of information about objects are represented by cortical activity patterns across the human visual pathway, ranging from low-level visual areas that respond best to basic features, to high-level areas that respond best to complex objects. Rather than focusing exclusively on the high-level object-selective areas, this project emphasizes a different approach to understand how invariant representations of objects are formed. Studies will characterize the neural representation of objects at each stage of the visual pathway, from the primary visual cortex to anterior inferotemporal areas, to determine how object representations are transformed from one processing stage to the next. This research will help reveal how the brain solves the problem of object recognition, by transforming the raw retinal input into increasingly more flexible representations of the object through a process spanning many successive levels of the visual pathway. Results from these studies will provide inform current theories of object recognition. Understanding these neural bases is necessary to comprehend what can go wrong with object recognition in cases of learning disability, developmental disorder or brain injury (e.g., developmental or acquired dyslexia). It is likely that computer algorithms for recognizing objects will also be improved.