April 14, 2017 -
3:15pm to 4:30pm
Jordan Hall (Building 420), Room 050
Chengxu Zhuang Title: Exploring goal-driven deep neural network models of sensory cortex
Chengxu Zhuang Abstract: Goal-driven deep neural network modeling methods have recently made progress in explaining neural response patterns in higher visual and auditory cortex. In this work, I will describe my work extending these ideas, both going broader into new sensory domains and going deeper into the visual domain. I will first describe my work modeling rodent somatosensory cortex. Mice use whiskers actively to sense their world. We hypothesize that neurons in mouse somatosensory cortex help it to estimate the shape and category of objects in the environment, invariant to size, speed, and angle of whisker attack. By building a virtual whisker array, using the known mechanical properties of mouse whiskers and stimulating those whiskers using various objects in different scales and orientations, we build a large dataset that can support the training of networks predicting object identity and surface shape using stimulated responses from whiskers touching the objects. I will describe our efforts to train a variety of deep neural network to achieve these tasks, expressing different hypotheses about the architecture of somatosensory cortex. If time allows, I’ll also discuss my work making better models of the primate visual system, by using rich intermediate tasks, such as depth perception and surface normals reconstruction, to both complement and supplement categorization task.
Department of Psychology