时间：2026年11月5日（周四）中午 13:00
地点：清华大学吕大龙楼11层1100
报告主题：To see and not to do is not yet to see: evidence from a mouse digital brain performing an ecological navigation task
报告人：博士研究生：殷子
内 容：
Can intelligence arise from observing the world, or does it also depend on actively interacting with it? Recent AI suggests that observation goes a long way: models trained on large collections of images and videos now match humans at object recognition and generate realistic video. Biology suggests otherwise. In Held and Hein's classic experiment (1963), kittens that received comparable visual input, but not through their own movement, failed to develop normal visually guided behavior. One reason for this divergence may be that the two kinds of vision serve different goals: artificial vision is built for representation, whereas biological vision serves control.Testing this idea requires a complete sensorimotor loop, which is hard to observe in the brain and largely missing from existing models. We therefore built a virtual mouse that navigates a maze from a first-person view: a data-constrained model of mouse primary visual cortex (V1) drives a controller in real time, and the mouse's own actions determine what it sees next. With this system we ask three questions. First, what visual information guides navigation? Separating V1 activity into a geometry stream and a motion stream, we find that the motion stream, which keeps direction but discards position, is critical for turning. Second, how does biological vision compare with pretrained AI vision? On the training routes, pretrained vision models perform even better than V1. On unseen routes, they fail at every critical turn, whereas the V1 system still turns correctly at nearly half of them, and this ability comes from the motion stream. Third, can vision support downstream spatial representation? With V1 input alone, a recurrent network develops grid-like units, and the motion stream favors this periodic structure. Together, these results suggest that observation alone may not be enough: vision shaped for control provides priors that support generalization, pointing to a shift from cloud AI toward physical AI.
来源：刘嘉课题组
排版：杨蓝瑜
责编：牛政旸
审核：伍珍
