Can Winner-Take-All Mechanism Underlie Pop-Out Visual Search Computation?
by Mr. Ori Hendler
Dept. of Physics and Zlotowski Centre for Neuroscience, Ben-Gurion University of The Negev
at Biological and soft-matter physics
Thu, 13 Jan 2022, 12:10
ZOOM only - Meeting ID: 820 9132 7859 Passcode: 345123
A prominent theory in the study of pop-out visual search is the proposition that the underlying neuronal computation is based on a set of contextually modulated cells that process the entire visual scene simultaneously followed by a Winner-Take-All competition that reads out the information. Yet, an inherent issue is that past studies of Winner-Take-All readout have showed that it is a highly sensitive mechanism. Specifically, the readout, performance of Winner-Take-All is very poor in its ability to accumulate information from large populations. These observations raise the question whether a Winner-Take-All can reliably identify a deviant stimulus in the background of numerous distractors. To address this question, we investigated the performance of a Winner-Take-All readout in a pop-out task. This was done using the framework of a modeling study which combines real data from several animal species together with analytic and numeric evaluation of the error. First, we studied Winner-Take-All performance in a competition between one deviant and many distractors homogenous populations. We found that performance of the Winner-Take-All deteriorates rapidly with the number of distractors is increased while improving very slowly with the number of neurons in each population. Moreover, introducing into the model the inherent neuronal heterogeneity prevents the Winner-Take-All from reading out information from large neuronal populations. Overall, our results demonstrate that it is unclear whether Winner-Take-All can account alone for the rapid computation of salient features or the correct identification of a pop-out stimulus.
Authors: Ori Hendler, Ronen Segev, Maoz Shamir
Join Zoom Meeting
Meeting ID: 820 9132 7859
Created on 09-01-2022 by Granek, Rony (rgranek)
Updaded on 09-01-2022 by Granek, Rony (rgranek)