Neuroscience 2005 Abstract
| Presentation Number: | 287.14 |
|---|---|
| Abstract Title: | A recurrent neural network model of the temporal dynamics of spatial remapping. |
| Authors: |
Keith, G. P.*1,2
; Wang, H.2
; Crawford, J. D.1,2,3
1Psychology, York Univ., Toronto, Canada 2Centre for Vision Research, York Univ., Toronto, Canada 3Biology, York Univ., Toronto, Canada |
| Primary Theme and Topics |
Sensory and Motor Systems - Visuomotor Processing -- Spatial memory and sensorimotor transformations |
| Secondary Theme and Topics | Sensory and Motor Systems<br />- Eye Movements<br />-- Saccades and Pursuit |
| Session: |
287. Visuomotor Processing: Spatial Memory and Sensorimotor Transformations I Poster |
| Presentation Time: | Sunday, November 13, 2005 2:00 PM-3:00 PM |
| Location: | Washington Convention Center - Hall A-C, Board # Z26 |
| Keywords: | MODELING, OCULOMOTOR, SUPERIOR COLLICULUS, BRAIN STEM |
In a previous study (Keith, Smith & Crawford, sfn Abstracts 2004) we trained a 3-layer feed-forward neural network using back-propagation to perform the remapping of visual target positions across an intervening saccade. Units in the hidden layer showed two properties seen in LIP and deeper layers of the superior colliculus (SC): predictive remapping and open-ended receptive fields. In the current study we added dynamic elements, first in the form of recurrent connections between all hidden-layer units, with a one-time-step delay associated with this signal. The inputs to the network were the initial retinal error of the visual target, represented as a topographic array of units, and eye position and motor error, both represented by a push-pull pairs of units for each degree of freedom. The output of the network was the dynamic motor error of the saccade to the target. The network was trained using back-propagation-through-time to perform the remapping of the target across the intervening saccade. The network sustained the output activation in time-steps after the inputs were discontinued, showing the latching property characteristic of LIP neurons. We then fed the network output motor error back into the network as the input motor error as a dynamic loop. The network drove the output topographic distribution of activation to the zero motor error position, showing the moving hill of activation seen in the SC. When the feedback signal was suspended, the output of the network ceased its movement. It continued this movement when the feedback was resumed, again an observed property of the SC. This study thus shows how several properties consistent with those seen in saccade-associated brain areas may be produced spontaneously using very simple recurrent neural networks, without imposing additional constraints on the implementation.
Supported by NSERC and OGS Canada
Sample Citation:
[Authors]. [Abstract Title]. Program No. XXX.XX. 2005 Neuroscience Meeting Planner. Washington, DC: Society for Neuroscience, 2005. Online.
Copyright © 2005-2026 Society for Neuroscience; all rights reserved. Permission to republish any abstract or part of any abstract in any form must be obtained in writing by SfN office prior to publication.
