Abstract
Change blindness experiments suggest that we do not construct a fully elaborated representation of the external world. I present here a novel computational model to explain how higher cognitive processes influence early visual processing, in order to access the external world. Notably, within this model visual attention is not explicitly implemented, but rather emerges as a natural consequence of competitive interactions between brain regions (specifically, in this case, V4, IT, FEF and PFC).
It is proposed that visual areas first acquire information about the scene by processing visual stimuli in a parallel bottom-up manner and acquire a more detailed knowledge about an object of interest by feedback. Activity from the ventral pathway then converges onto premotor neurons to enforce behavioral decisions (in this case FEF movement cells), which in turn are fed back onto cells in the ventral pathway. This reentrant activity has the effect of implementing a spatially organized gain control of the item of interest.
This model successfully replicates previously collected data from single-cell recordings in IT and V4 from a macaque visual search task (Chelazzi, Duncan, Miller, Desimone, J. Neurophysiol. 1998; Chelazzi et al., Cereb Cortex 2001). It predicts that the late target discrimination around 150ms after scene onset, often identified with conscious processing, is based on reentry from premotor neurons. The model is also consistent with other recordings in FEF (Schall, Philos Trans R Soc Lond B Biol Sci 2002). Moreover, by representing visual features as populations of active cells, whose activity is dynamically modulated in a parallel manner, this model offers an efficient solution to the problem of target localization within a cluttered visual display (e.g. natural scenes). Attention is not a prerequisite for object recognition but the dynamics resolve the ambiguities within active populations.