RT Journal Article
A1 White, Brian
A1 Berg, David
A1 Ikeda, Takuro
A1 Levy, Ron
A1 Itti, Laurent
A1 Munoz, Douglas
T1 Comparison of superior colliculus and primary visual cortex in the coding of visual saliency
JF Journal of Vision
JO Journal of Vision
YR 2014
DO 10.1167/14.10.517
VO 14
IS 10
SP 517
OP 517
SN 1534-7362
AB  The superior colliculus (SC) is a phylogenetically ancient midbrain structure with purely visual representations in the superficial-layers (SCs), and sensorimotor representations linked to the control of eye movements/attention in the intermediate-layers (SCi). In primitive species, SC played a central role in vision and orienting independent of cortex. Through mammalian evolution, primary visual cortex (V1) introduced more specialized feature processing, but also became a dominant source of input for SCs. We quantified the relative roles of these early visual areas in the coding of higher-order stimuli that give rise to saliency in complex scenes. Rhesus monkeys viewed a wide-field arrangement of stimuli (210 radially-arranged items spanning ~40-50deg) extending beyond the classic receptive field (RF). The stimuli were oriented color bars (~0.4x1.2deg) that formed a perceptual "pop-out" array the monkeys had to ignore; i.e., reward was contingent upon gaze directed to a separate achromatic stimulus that always stepped orthogonal to the pop-out singleton. We compared visually-evoked responses when the goal-irrelevant pop-out singleton fell in versus opposite the RF, and compared it to a single-item control condition. First, visual onset latency was reliably earlier, and RFs 5-10 fold smaller, in V1 than SCs (or SCi). This is consistent with the idea that each SCs neuron might integrate the outputs of multiple V1 neurons. Second, surround suppression evoked by the wide-field array was dramatically stronger in SCs (and SCi) than V1. This indicates the prominence of long-range interactions in SC, an essential component of the saliency-map hypothesis. Lastly, only SCs neurons showed a reliable preference for the goal-irrelevant pop-out singleton; in SCi stimuli were heavily suppressed unless made goal-relevant. From these observations, we propose that V1 fits the role of a local feature processer, SCs a bottom-up saliency map, and SCi a priority map shaped by a combined representation of bottom-up saliency and top-down relevancy. Meeting abstract presented at VSS 2014 
RD 4/20/2021
UL https://doi.org/10.1167/14.10.517