Technical Program

Paper Detail

Paper: PS-1A.23
Session: Poster Session 1A
Location: Symphony/Overture
Session Time: Thursday, September 6, 16:30 - 18:30
Presentation Time:Thursday, September 6, 16:30 - 18:30
Presentation: Poster
Publication: 2018 Conference on Cognitive Computational Neuroscience, 5-8 September 2018, Philadelphia, Pennsylvania
Paper Title: Laterally connected neural field provides precise centroid estimates
Manuscript:  Click here to view manuscript
Authors: Sebastian Waz, Charles Chubb, University of California, Irvine, United States
Abstract: The convolution operator is an essential tool in digital signal processing. Among many other things, it grants tremendous perceptual ability to a popular class of deep learning models known as convolutional neural networks; however, the convolution operator's neural implementation typically requires dense connectivity, so it also diminishes the biological plausibility of such models. We found that the convolution of an input vector with a parabolic function can be performed without dense connectivity. Specifically, this operation can be performed by a laterally connected (i.e., biologically plausible) neural network that evolves in continuous time (a ``neural field''). This particular convolution is shown to have useful properties for centroid estimation. Rapid and precise centroid estimation is known to take place early in the human visual system, but to date, this computation lacks an adequate biologically plausible neural model. Using the aforementioned results, we present such a model.