Computational foundations of active visual sensing

点击翻译按钮获取中文摘要

Abstract Vision is an active process: we move our head and eyes to explore the sensory world. This is particularly important in situations where a stationary view provides limited information, such as when looking for an object that is occluded or obscured, which is common in complex natural scenes. However, our understanding of active vision is limited due to experimental and theoretical challenges, including the difficulty of studying vision in freely moving animals and the lack of formal theoretical frameworks that integrate visual representations with actions. In this team project, we will combine expertise in visual neuroscience, behavior, machine learning, and theory, to determine the behavioral, neural, and computational underpinnings of active sensing. Our approach is based on a new theoretical framework of Bounded Rational Control (BRC), and a behavioral task in which mice perform an object recognition task in the presence of occlusion and image corruptions. To enable active sensing, stimuli in the task are rendered real-time in augmented reality based on the animal's viewpoint. In our first aim, we will develop models of active sensing based on constrained visual representations in BRC. In the second aim, we measure behavioral performance (both correct/incorrect responses and full-body movements) during the task, and in the third aim we will measure neural activity across visual cortical areas during the task. For both Aims 2 and 3, we will fit our models to the corresponding behavioral and neural data, and then perform causal tests of our models by presenting novel stimuli predicted to elicit specific responses from the model. Together, these aims will provide a foundational understanding of active vision in the mouse that will support a subsequent U19 proposal taking advantage of genetic tools to investigate the underlying local and long-range neural circuits.