Technical Program

Paper Detail

Paper: PS-2A.8
Session: Poster Session 2A
Location: Symphony/Overture
Session Time: Friday, September 7, 17:15 - 19:15
Presentation Time:Friday, September 7, 17:15 - 19:15
Presentation: Poster
Paper Title: Unsupervised learning of manifold models for neural coding of physical transformations in the ventral visual pathway
Manuscript:  Click here to view manuscript
Authors: Marissa Connor, Christopher Rozell, Georgia Institute of Technology, United States
Abstract: Biological vision is envied for its ability to learn to recognize objects in the 3D world undergoing physical transformations. A recent hypothesis is that the ventral visual pathway exploits the manifold nature of these transformations to form neural codes that are efficient for discrimination, but there is no compelling model for how this representation is learned from data. We propose a computational model that performs unsupervised learning on received retinal imagery to infer identity-preserving transformations. We show that such a model can successfully learn useful representations on a subset of objects that can be transferred to new objects. We also demonstrate that this model can be used to infer 3D transformations from 2D imagery despite the ill-conditioned nature of the problem. This model for 3D inference can account for psychophysical experiments such as 3D shape perception from random-dot kinematograms.