Abstract
Visual attention is often understood as a modulatory field at early stages of processing. In primates, attentive selection is influenced by figure-ground segregation, which occurs at early stages in the visual cortex. The mechanism that directs and fits the field to the object to be attended is not known. We propose here that the same neural structures that serve figure-ground organization automatically focus attention onto a perceptual object. Specifically, we show that an additive attentional input which is spatially broad and not tuned for object scale produces a quasi-multiplicative attentional modulation which is repositioned and sharpened to match the object contours (auto-localization) and tuned for the scale of the object (auto-zoom). The model quantitatively reproduces the changes in attentional modulation caused by the presence of objects observed at the level of V2. The proposed mechanism works with generic, zero-threshold linear neurons, additive inputs and the connection patterns are plausibly related to the statistics of natural visual scenes. We performed a global sensitivity analysis to determine the dependence of the attentional modulation, border ownership modulation and their interaction on several parameters in the model. The pattern and strength of the lateral inhibition are key to obtaining a sharpening of the attention field and a quasi-multiplicative attention modulation with an additive attention input. The strength of reciprocal connections from neurons representing local features and neurons integrating them, and inhibition between inconsistent proto-object representations are important to repositioning and tuning for scale of the attention field.
5R01EY016281-02 and R01-NS40596.