FriSem

Date
Fri April 12th 2019, 3:15 - 4:30pm
Event Sponsor
Department of Psychology
Location
Jordan Hall Room 050

Dan Bear, Postdoc with Prof. Daniel Yamins, Department of Psychology, Stanford University

Title: "How Should We Build an Artificial Visual System?"

Abstract:Vision is one of evolution's far-reaching inventions, helping us solve a wide range of problems posed by the environment. We want to understand how this works: what computations are carried out by the brain's visual system, and how can they be implemented in artificial machines?Research in computer vision has found one class of algorithm, the task-optimized Convolutional Neural Network (CNN), that captures some of our abilities and partly explains how visual information is represented by real neurons. But CNNs aren't complete models of vision. The 2D-image-like internal representations of CNNs do not allow for the temporally-unfolding computations of real neural populations, and they are ill-suited for solving problems in 3D, physical environments.I will explain how different artificial neural network structures could bring computer vision closer to our own. In particular, certain recurrent network architectures can explain population dynamics in the ventral visual system and may yield better models of how we recognize objects. For tasks beyond object recognition, a new type of internal representation -- which explicitly encodes the discrete, physical properties of 3D scenes -- can be used by non-CNN algorithms to predict future scene dynamics. I will show work-in-progress on how to construct this new representation from visual input.