Related Experiment Videos
Self-organizing task modules and explicit coordinate systems in a neural network model for 3-D saccades.
1Centre for Vision Research, and Department of Psychology, York University, Toronto, Ontario, Canada. mas@yorku.ca
Journal of Computational Neuroscience
|May 22, 2001
Summary
This study trained an artificial neural network to generate accurate eye movements (saccades). The network developed specialized internal modules, including a coordinate system, to perform visuomotor transformations, mimicking biological systems.
Area of Science:
- Computational Neuroscience
- Artificial Intelligence
- Ophthalmology
Background:
- Understanding the neural mechanisms of eye movements (saccades) is crucial for neuroscience.
- Visuomotor transformations involve converting visual information into motor commands for eye control.
- Artificial neural networks offer a powerful tool for modeling complex biological processes.
Purpose of the Study:
- To train an artificial neural network to generate accurate saccades in Listing's plane.
- To analyze how the hidden units within the network perform the visuomotor transformation.
- To investigate if error-driven learning can produce functional modules and coordinate systems similar to biological saccade generators.
Main Methods:
- A three-layer artificial neural network was trained using back-propagation.
- The network received oculocentric retinal error vectors and 3D eye orientation as input.
- The network's hidden layer activity was analyzed to understand the visuomotor transformation process.
- Selective 'lesions' were performed on network modules to assess their functional contribution.
Main Results:
- The trained network successfully generated accurate head-centric motor error vectors for saccades within Listing's plane.
- Hidden units did not explicitly represent target direction or eye orientation.
- Units self-organized into four parallel modules, including a dominant 'vector-propagation' class.
- The 'vector-propagation' module formed a precise orthogonal coordinate system aligned with Listing's plane.
- This module primarily drove saccade magnitude and direction, while other modules modulated eye position.
Conclusions:
- Error-driven learning is sufficient to create discrete functional modules and explicit coordinate systems in artificial neural networks.
- The emergent properties of the trained network closely resemble those observed in biological saccade generators.
- This study provides insights into the computational principles underlying visuomotor control.