A dynamical neural simulation mapping feature-based attention to location with non-linear cortical circuits

Abstract

Visual attention mechanisms allow the visual cortex todirect cortical processing to salient locations within avisual scene, at the expense of diminished processing ofstimuli at other locations. The deployment of attentionto a given location is known as spatial attention.Attention may also be applied to a known feature in avisual scene, such as a colour or shape, enhancing itsoccurrences across the visual field. In [1], a model ofbinding distributed feature representations throughfeature-based attention predicted that attentional feed-back would act in all positions of the visual scene in lowlevel visual areas. There is now experimental evidencethat this mechanism exists [2].We present here a neural dynamical model of feature-based attention that directs attention to salient locationsinamulti-objectscenethrough interaction of stimulusdriven and cue driven neural activation. This interactionis managed by a disinhibition circuit gating output to aretinotopic neural region of the dorsal stream, thelateral intraparietal area (LIP). When visual stimuli arepresented to primary visual cortex (V1), neural activa-tion spreads into higher ventral stream areas throughincreasing neural receptive fields between layers.Stimulus driven neurons activate excitatory output andinhibitory interneuron populations in the disinhbitioncircuit. When a salient feature is cued by activation ofits associated anterior inferotemporal (AIT) neuronalpopulation, neural activity propagates to lower visualareas across all locations of the visual field. Activity inthis top-down flow inhibits the inhibitory interneuronpopulation allowing excitatory output from the disinhi-bition circuit to LIP in a retinotopic manner whereactivity is correlated in the stimulus and cue drivenpathways. In areas where the two pathways do notmatch, local inhibition in the disinhibition layer preventsneighbouring circuits’ excitatory populations from out-putting to LIP.The model exhibits effects of feature-based attention,such as the feature similarity gain principle and biasedcompetition through interactions of stimulus and cueinitiated neural activity. The dynamical nature of themodel as neural populations of Wilson-Cowan differen-tial equations allows the time-course of neural activityto be investigated.ConclusionsAlthough originally designed as a model for binding, themodel replicates the experimental correlates of feature-based attention, which strongly suggests its functionalrole is actually binding An important prediction of themodel is that local non-matches between cue andstimulus are the most important information projectedfrom the ventral to the dorsal stream and that this inter-action is highly non-linear. We present candidates forneural circuits implementing this interaction.