Computational Models and Physiological Studies of Feedback in Visual Object Recognition Tasks

对我们周围的物体、场景和人的视觉识别依赖于视觉皮质，这是灵长类大脑中最大和最复杂的部分之一。这一区域也是认知科学中的一个主要难题：大脑皮层反投影的作用是什么，也就是说，从次级区域返回到初级区域的联系？在这个项目中，Tomaso Poggio博士和Earl Miller博士试图解释复杂视觉识别任务背后的神经计算。他们的研究将有助于将从视觉感知动物模型中获得的知识转移到对人类高级认知视觉过程的理解上。灵长类物体识别系统的最新计算理论与不同视觉区域的各种生理发现相一致，例如V1(初级视觉皮质)、V4(视觉区域4)、IT(颞下皮质)和PFC(前额叶皮质)。更令人惊讶的是，它模仿了人类对复杂自然图像进行快速分类的行为表现，并在困难的识别任务中表现出与几个最先进的计算机视觉系统一样的性能。考虑到该模型能够解释快速物体识别，而且它目前只使用前馈处理，一个重要的谜题涉及到已知存在于大脑皮质的丰富解剖背投影的计算和功能作用。拟议的项目采取了双管齐下的方法来寻找它们的功能。首先，腹侧流计算模型的实验将为理解反投影的可能功能提供有用的见解。其次，在猕猴身上进行的多单位记录实验将在IT和PFC水平上表征几种不同识别任务中的自上而下效应及其时间。这项分析将利用最近开发的一项自动分类技术，该技术将大脑活动与单个视觉对象联系起来。通过结合模型和实验的结果，可以检验对反馈加工的认知作用的计算解释。了解视觉中背向投影的功能将有助于我们理解视觉本身的神经基础，但它也将帮助我们理解大脑的整体设计。这个器官的复杂性继续让我们感到惊讶，通过像视觉这样相对容易理解的领域来掌握它的机制，是扩大我们对整个系统的欣赏的一个有希望的途径。另一个目标是证明计算理论，特别是视觉理论和实验之间的相互作用，将使理解大脑功能变得更容易。这些进步帮助我们理解大脑的正常功能，并允许我们在大脑不正常工作时提供更好的帮助，让机器看得更清楚，并为机器人技术带来新的方法。

Visual recognition of objects, scenes and people around us depends on the visual cortex, one of the largest and most complex parts of the primate brain. This area also presents one of the major puzzles in cognitive sciences: What is the role of the cortical back-projections, that is, links from the secondary areas back into the primary areas? In this project, Drs. Tomaso Poggio and Earl Miller try to interpret the neural computations underlying complex visual recognition tasks. Their studies will help to transfer knowledge gained from animal models of visual perception toward the understanding of higher cognitive visual processes in humans. A very recent computational theory of the primate object recognition system agrees with a variety of physiological findings in different visual areas, such as V1 (primary visual cortex), V4 (visual area 4), IT (inferotemporal cortex), and PFC (prefrontal cortex). Even more surprisingly, it mimics human behavioral performance on rapid categorization of complex natural images, and performs, as well as several state-of-the-art computer vision systems on difficult recognition tasks. Considering that the model is able to account for rapid object recognition, and that it currently only uses feedforward processing, a significant puzzle concerns the computational and functional role of the abundant anatomical back-projections known to exist in cortex. The proposed project takes a two-pronged approach toward finding their function. First, experiments with the computational model of the ventral stream will provide useful insights for understanding the possible functions of the back-projections. Second, experiments with multi-unit recordings in macaques will characterize top-down effects and their timing in several different recognition tasks at the level of IT and PFC. The analysis will make use of a recently developed automatic classification technique that relates brain activity with individual visual objects. By combining results from the modeling and experimentation, computational explanations for the cognitive role of the feedback processing can be tested. Understanding the function of back-projections in vision will help us understand the neural basis of vision itself, but it will also help us understand the global design of the brain. The intricacy of this organ continues to amaze us, and getting a handle on its mechanism through a relatively well-understood area like vision is a promising avenue for expanding our appreciation of the whole system. A further goal is to show that the interaction between computational theories, in particular of vision, and experiments will make it easier to comprehend brain functions. Such advances help us appreciate the normal function of the brain and allow us better means of helping when it does not function normally, of making machines see better, and of bringing new approaches to robotics.