An Image-Computable Model of Orientation-Tuned Normalization

Divisive normalization has been proposed to be a canonical neural computation, describing how a population of cortical neurons mutually inhibit one another. Even in V1 however, where the model was initially applied, it remains uncertain what comprises the normalization pool of a given neuron. Does normalization depend on all nearby neurons (untuned normalization), such that responses are effectively normalized by local stimulus contrast? Or is normalization selective, for example where neurons tuned to similar features are weighted more strongly (tuned normalization)? In the last decade, psychophysical, electrophysiological, and fMRI studies suggested that normalization depends on feature similarity, and therefore on statistical regularities within images. We implemented an image-computable, normalized energy model in which normalization is contingent on matched orientation tuning. The model was fit to human electrocorticographic (ECoG) data from 7 pre-surgical patient volunteers with implanted electrodes on the surface of visual cortex. Participants viewed a range of gray-scale, band-pass filtered, static images (96 to 116 unique images) for 500 ms each. For each electrode and image, we computed the broadband power (50-200Hz) during stimulus presentation. We identified 21 electrodes over V1 with population receptive fields centered within the stimulus aperture (up to 10° eccentricity). Contrast energy was extracted from each image using a steerable pyramid, yielding 42 values at each image pixel (6 orientation x 7 spatial frequency channels). Each output was divisively normalized, with the normalization pool composed of nearby spatial locations, all spatial frequencies, and matched orientation. These outputs were then summed across orientation and spatial frequency channels within a 2D Gaussian spatial pRF. Model accuracy was high (~70%), considerably higher than comparison models with untuned normalization or without normalization. Our results demonstrate that visual responses are best captured when the normalization pool incorporates feature similarity, a finding that can be applicable to many experiments and datasets.