Fig. 6: Surround suppression or facilitation of V1 correlations is predicted by the optimal pairwise GSM structure for natural images.
A We binned model neuron pairs by their signal correlations and RF distances (center-to-center distance), and computed the average log-likelihood ratio per bin for shared versus independent pairwise GSM across 10,000 natural images from the ImageNet test set (see Methods for details). The green (purple) entries indicate conditions where the shared (independent) GSM model is the better model of image statistics on average. We observe a sharp transition at specific RF distances, where the overlap between the centers and surrounds of two model neuron filters is reduced. B We binned model neuron pairs as in (A) and computed the modulation of correlations per bin for 500 natural images, i.e., the difference in rsc for small images minus large images. For pairs with overlapping receptive fields and high tuning similarity, correlations are often suppressed (blue). In contrast, for pairs with non-overlapping receptive fields and low tuning similarity, correlations are often facilitated (red). C We binned V1 neural pairs and computed the modulation of correlations per bin as in panel (B). Only pairs with at least one neuron RF centered on the stimulus were included (see Methods). Correlations were on average suppressed when neurons RFs are separated by less than 1° (blue) and facilitated otherwise (red), consistent with the model prediction.
