Download "Visual Saliency Computations: Mechanisms, Constraints, and the Effect of Feedback"

The primate visual system continuously selects spatial proscribed regions, features or objects for further processing. These selection
mechanisms— collectively termed selective visual attention—are guided by intrinsic, bottom-up and by task-dependent, top-down
signals. While much psychophysical research has shown that overt and covert attention is partially allocated based on saliency-driven
exogenous signals, it is unclear howthis is accomplished atthe neuronal level. Recent electrophysiological experiments in monkeys point
to the gradual emergence of saliency signals when ascending the dorsal visual stream and to the influence of top-down attention on these
signals. To elucidate the neural mechanisms underlying these observations, we construct a biologically plausible network of spiking
neurons to simulate the formation of saliency signals in different cortical areas. We find that saliency signals are rapidly generated
through lateral excitation and inhibition in successive layers of neural populations selective to a single feature. These signals can be
improved by feedback from a higher cortical area that represents a saliency map. In addition, we show how top-down attention can affect
the saliency signals by disrupting this feedback through its action on the saliency map. While we find that saliency computations require
dominant slow NMDA currents,the signal rapidly emergesfrom successive regions ofthe network. In conclusion, using a detailed spiking
network model we find biophysical mechanisms and limitations of saliency computations which can be tested experimentally.

Leave a Reply