We made single-unit recordings in area V2 in two fixating rhesus macaques. We presented stimuli composed of multiple superimposed grating patches that localize contrast energy in space, orientation, and scale. V2 activity is modeled via a two-layer linear-nonlinear network, optimized to use a sparse combination of V1-like outputs to account for observed activity.

Analysis of model fits reveals V2 neurons to be well-matched to natural images, with units combining V1 afferent tuning dimensions to effectively capture natural scene variation. Remarkably, although the models are trained on responses to synthetic stimuli, they can predict responses to novel image classes, i.e. naturalistic texture, reproducing single-unit selectivity for higher-order image statistics. Thus, we demonstrate state-of-the art performance of modeling V2 selectivity, and provide a mechanistic account of single-unit tuning for higher-order natural features.