Recently, we have shown that by computing differences between populations of V1-like units, V2-like units can become sensitive to higher-order statistics of natural texture, beyond the oriented energy (spectral) features relayed by V1.

Here we test our theoretical predictions against single-unit recordings in areas V1 and V2 in an awake and fixating macaque. After characterizing a unit's receptive field using standard methods, we presented novel stimuli generated by superimposing patches of oriented gratings at multiple positions and scales. We fit a two-stage convolutional linear-nonlinear model to these responses: Stimuli are initially processed with a convolutional bank of V1-like filters selective for position, orientation, and scale. We then jointly optimize a linear-nonlinear combination of these rectified units to explain observed neural data. The model often explains a substantial fraction of response variance, comparable or superior to existing models. Qualitative differences emerge between models trained on V1 and V2 data. As expected, V1 models respond to a narrow range of position, orientation, and scale, whereas V2 cells often exhibit sensitivity to differences over stimulus position, scale, and orientation energy. By explicitly combining V1-like afferent activity, our two-stage model can explain V2's selectivity for higher-order stimulus features.