A distributed computational model of spatial memory anticipation during a visual search task¶

The objective of the article was to propose a model based on the dynamic neural fields computational paradigm that must be able to focus successively on several targets. As soon as a target is selected, the system must perform an eye movement or camera movement to center it on its retina.

There are five maps :

• Visual : The input image is filtered and clamped in that map. We take care of the eye movements to switch on each target. After each eye movement, the visual scene is moved, this is why the activity in the visual map is translated from what is drawn on its right, where the visual scene is supposed static.
• Focus : A competition between the neurons in that map, with locally excitatory and widely inhibitory pattern of connection, leads to only one target. This is the selected target.
• wm (working memory) and thal_wm : This circuit consists of 2 maps that are recurrently connected; this is the memory circuit. The basic condition for a stimulus to enter in that map is to be both in visual and focus. It means that a target is memorized as soon as it is a target we perceive and a target we focus on.
• Anticipation : that map is used to update the working memory after each eye movement, since after each eye movement, the positions of the targets we have memorized will change. The trick is that their new position are predictable.

In the simulation below, three targets are presented to the input, with an additional random noise. To observe the successive scanning of the three targets, you should first have a look to the activities in the input map. The “selected” target is the target emerging in the focus map. The co-activation of the focus and input map produces a sufficiently strong excitation to the working memory circuit (reverberating circuit between “working memory” and “thal_wm”) which induces the emergence of the selected target in working memory. The combined excitation of the working memory and the focus map leads to the anticipatory responses in the anticipation map, through sigmapi connections.

A saccade is manually triggered, though the dynamic of the system could be adjusted so that a saccade could be executed when the activities in the focus map reach a threshold. After the execution of a saccade, it is the co-activation of the anticipation and input maps that trigger the update of the working memory. An additional inhibitory bias between the working memory and the focus map prevents selecting a previously focused target.

On the demonstration below, only the order to execute a saccade is not automated. The C++ code required to run this simulation is available here code.tar.gz. This code was developed with Qt4.5.

You should wait that the animation has looped at least one time to see it fluently. A video is also available below the animation.