A Distributed Model of Spatial Visual Attention

Abstract

Abstract. Although biomimetic autonomous robotics relies on the massively parallel architecture of the brain, a key issue for designers is to temporally organize behaviour. The distributed representation of the sensory information has to be coherently processed to generate relevant actions. In the visuomotor domain, we propose here a model of visual exploration of a scene by the means of localized computations in neural populations whose architecture allows the emergence of a coherent behaviour of sequential scanning of salient stimuli. It has been implemented on a real robotic platform exploring a moving and noisy scene including several identical targets.

Objective validation of a dynamical and plausible computational model of visual attention

“…This solution has been partially developed through connectionist centralized models [11,12], actually they combined a hierarchical approach to produce a saliency map and a distributed model to generate the focus of attention. One of the main limitations is that competition between features (color, intensity and orientation) is done by the hierarchical part of the model.…”

Visual Attention for Robotic Cognition: A Survey

“…The neural network-based overt model reported in [77] is tightly coupled with biology (with respect to motor aspects of attention) and is focused on implementing visual exploration behavior guided by the novelty preference characteristics of primates attention. The identification of novelty in [77], however, is achieved though the implementation of space-based IOR, i.e., the robot moves to novel locations (through successive application of space-based IOR) and thereby attends to novel objects.…”

Cortical basis of communication: Local computation, coordination, attention

“…This mechanism is known as the inhibition of return [26] and is implemented with a couple of maps called Working memory in Figure 1 and described in [34]. One map in the Working memory receives topological (one to one) excitations from the Focus and the Saliency maps and sends reciprocal excitatory topological connections to another similar map that helps to maintain excitation: both external inputs are necessary to create excitation but only one (from the Saliency map) is sufficient to maintain it.…”

“…An initial model [34] including the Focus and Working memory maps is able to perform covert attention and displays similar properties [10]: if the desired target is encoded only in a feature map, it can directly bias the competition in the Focus map; else a selection corresponding to only one feature dimension is made, must be compared to the desired target and eliminated (thanks to the inhibition of return) or selected (hence a search time proportional to the number of targets).…”