Abstract
The method of population receptive fields (pRFs) allows quantitative modeling of brain activity responses to external stimuli, and has been used extensively to investigate cortical organization in health and disease (Dumoulin & Wandell, 2008, Dumoulin & Knapen, 2018). Since its introduction, the method has been extended to capture suppressive surrounds and compressive spatial summation (Zuiderbaan et al, 2012; Kay et al, 2013). Interestingly, suppressive surrounds explain more signal variance in early, but not late visual cortex; whereas the converse is true for compressive spatial summation. Divisive normalization has been proposed to underlie many psychophysical and neurological phenomena, and is a prime candidate for a canonical neural computation (Carandini & Heeger, 2012). Here, we build a pRF model based on divisive normalization and ask whether it can 1) unify previous pRF models and 2) provide new insights into cortical computations.
We acquired ultra-high-resolution 7-Tesla BOLD-fMRI data (1.7mm isotropic, 1.5s TR), while participants viewed high-contrast checkerboard bars passing on a mean-luminance background. We systematically varied the spatiotemporal properties (bar speed and width) of the stimulus. Participants performed a dot-color-discrimination task at fixation.
We show that divisive normalization pRFs parsimoniously explain both suppressive surrounds and compressive spatial summation, at a level consistently equivalent or above existing models. Furthermore, the divisive normalization pRF model includes new parameters that vary in biologically relevant ways, and lend themselves to interpretation in terms of neural baseline activity. For instance, we find that neural baseline activity varies systematically as a function of eccentricity.
Divisive normalization pRFs provide a biologically-inspired, unified modeling framework for seemingly different properties of responses to spatial visual stimuli observed across the visual hierarchy. In addition to unifying previous pRF models, divisive normalization allows quantification of neural baseline activity. Estimating neural baseline activity has many potential applications in neuroscience, perception, attention, and clinical conditions.