A model of visual-spatial memory across saccades.

Area of Science:

• Computational Neuroscience
• Cognitive Neuroscience
• Artificial Intelligence

Background:

• Oculomotor control involves complex neural mechanisms for directing eye movements.
• Short-term memory is essential for maintaining spatial information of visual targets.
• Previous models have explored leaky integrator mechanisms for oculomotor integration.

Purpose of the Study:

• To develop a neural network model capable of memory-guided saccades.
• To investigate short-term spatial memory mechanisms in neural networks.
• To explore biologically plausible neural representations for saccade memory.

Main Methods:

• Training a recurrent neural network with a hidden layer to perform memory-guided saccades.
• Implementing a leaky integrator mechanism with fixed point attractors for value decay.
• Analyzing receptive field properties of hidden units under varying parameters.

Main Results:

• The trained network accurately maintains stored spatial locations for several seconds.
• The leaky integrator mechanism demonstrates robustness to input/output variations.
• Biologically plausible parameters yielded hidden unit behavior similar to real neurons involved in saccade memory.

Conclusions:

• The developed neural network effectively models memory-guided saccades.
• The leaky integrator mechanism provides a robust framework for short-term spatial memory.
• Simultaneous representation in multiple reference frames can yield biologically realistic neural properties.