US-German Research Proposal: Neurocomputation in the Visual Periphery: Experiments and Models

周边视觉占视野的99.99%以上。它的优点和局限性强烈地限制了视觉感知——人类一眼就能看到的东西，以及他们移动眼睛拼凑世界信息的过程。周边视觉与中央凹视觉在复杂而有趣的方面有所不同，最重要的是由于“拥挤”，在这种情况下，识别周边刺激可能会因附近其他刺激的存在而严重受损。该项目将通过开发和测试一套周边视觉模型来研究视觉皮层编码的本质。这些模型的目标是回答有关神经生物学机制的关键问题。来自美国和德国的合作研究人员将开发模型，并创建一个用于解释行为结果的基准数据集。这些模型和数据集将免费提供，以帮助其他研究人员，并为开发平视显示器和用户界面等应用程序提供信息。这项工作将深入了解哪些特征是在视觉皮层中编码的，以及哪些权衡可能导致视觉系统发展出这种编码。理解这些权衡可能会让计算机视觉像人类视觉一样面临处理能力的限制。新模型变体的开发将基于神经生理学、自然图像统计、稀疏编码和卷积神经网络最近在人工智能中的成功。研究人员将通过研究自然图像任务和模型驱动实验的结合，收集比现有拥挤数据集丰富得多的基准行为现象。然后，他们将在基准数据集上比较新模型的预测结果，以及罗森霍尔茨博士现有的高性能周边视觉模型。这样做将识别出性能最好的模型，并回答有关池化计算和非线性运算符的本质，以及由周边视觉编码的特征的复杂性、性质和目的的关键问题。德国联邦教育和研究部（BMBF）正在资助一个伙伴项目。

Peripheral vision comprises over 99.99% of the visual field. Its strengths and limitations strongly constrain visual perception -- what humans can see at a glance, and the processes by which they move their eyes to piece together information about the world. Peripheral vision differs from foveal vision in complex and interesting ways, most importantly due to "crowding," in which identifying a peripheral stimulus can be substantially impaired by the presence of other, nearby stimuli. This project will examine the nature of the encoding in visual cortex, through development and testing of a set of models of peripheral vision. These models will be targeted at answering key questions about the neurobiological mechanisms. The collaborating investigators, in the US and Germany, will develop models and create a benchmark dataset of behavioral results to be explained. The models and dataset will be made freely available, to aid other researchers and to inform the development of applications such as heads up displays and user interfaces. This work will provide insight into what features are encoded in visual cortex, as well as what tradeoffs may have led the visual system to develop that encoding. Understanding those tradeoffs may inform computer vision which, like human vision, faces constraints on processing capacity. The development of new model variants will be based on insights from neurophysiology, natural image statistics, sparse coding, and the recent success of convolutional neural networks in artificial intelligence. The investigators will gather benchmark behavioral phenomena far richer than existing crowding datasets, through a combination of studying natural image tasks and model-driven experiments. They will then compare predictions of the new models, as well as of Dr. Rosenholtz's existing high-performing model of peripheral vision, on the benchmark dataset. Doing so will identify the best-performing model(s), and answer key questions about the nature of pooling computations and of non-linear operators, and about the complexity, nature, and purpose of the features encoded by peripheral vision.A companion project is being funded by the Federal Ministry of Education and Research, Germany (BMBF).