Abstract
Context integration is a fundamental feature of the human visual system contributing to basic visual processing routines such as contour integration, visual pop-out, and texture segmentation. Two perceptual hallmarks of context integration are orientation-dependent surround suppression (ODSS) and the orientation tilt illusion. In these contextual effects, either the magnitude (ODSS) or the orientation (tilt illusion) of a neural population vector code is shifted depending on the relative orientation of the spatial surround. Supported by past modeling studies, these perceptual effects have been attributed to a divisive gain control mechanism, in which the spatial surround exerts an orientation-tuned modulatory influence on the center. Electrophysiological measures have shown that oscillations in the gamma band (30-100Hz) of the local field potential (LFP) in visual cortex may serve as a biomarker for gain control. However, previous models of context integration have not explained this aspect of neural dynamics in connection to gain control. To address this issue, we use oscillatory recurrent gated neural integrator circuits (ORGaNICs), a model based on a biological implementation of long short-term memory (LSTM) networks, that has been shown to generate oscillatory dynamics in V1. By constraining the model using a recurrent circuit motif consistent with primate neurophysiology, we explain both perceptual effects and oscillatory dynamics observed in ODSS. We show that the model is capable of matching both psychophysical measures of the tilt illusion and narrowband gamma power observed in LFPs recorded from human visual cortex. Furthermore, the model predicts that similarly oriented surround stimuli increase the strength of gain pools for the center causing an increase in gamma power, an experimentally testable prediction.