Figure 6. Performance of the complex-cell feedforward linear model.

(A) Responses of V2 cells/pixels to excluded grating stimuli of 8 different orientations predicted by the complex cell model in the leave-one-out (LOO) procedure (y axis) versus responses measured experimentally (x axis). This plot shows data for all V2 cells/pixels in eight injection cases. The mean correlation coefficient between measured and predicted responses across the 8 LOO tests is 0.88±0.04. (B) Averaged relative cross-validation error for each injection case. Zero indicates a (perfect) prediction, while a value of 1 indicates that the difference between predicted and measured response equals the range (i.e. max minus min) of the cell’s measured response to the full set of 8 orientations; in practice the relative errors were never >0.6. The inset shows a color-coded relative error map at the V2 injection site for a representative case (MK373-CTB647). (C) Averaged absolute error (defined as the absolute value of predicted PO minus measured PO) in the model’s prediction of the preferred orientation (PO) of V2 cells/pixels. The inset is a color-coded signed error map calculated for each pixel at the V2 injection site for case MK373-CTB647. (D) Averaged absolute error in the model’s prediction of the width of the tuning curves (HWHH: half width at half height) of V2 cells/pixels. The inset is a color-coded signed error map calculated as predicted HWHH minus measured HWHH for each pixel at the V2 injection site for case MK373-CTB647. Error bars: s.e.m.