A competitive network for multi-object selection

Psychophysical and neurophysiological investigations shows that visual system is able to simultaneously select and attend to multiple objects present in the visual scene. A neural network is proposed with the ability to select multiple objects simultaneously if they share the same level of saliency. The model is based on the segmentation network which labels all spatial locations occupied by the object with the same activity level. Computer simulations showed that the model can perform visual search task with parallel access to multiple instances of the same object. Visual search is implemented using dynamic routing circuit which achieves translation-invariant representation of selected objects. Attentional shifts are imple-mented using template matching between sensory and memory representation. If there is a mismatch between them, a global inhibition drives the whole network to search for a new pattern. The proposed network is able to select objects even when their visual representation is corrupted with Gaussian noise. The model´s behavior is consistent with the Boolean map theory of visual attention.