Fig. 7
Responses explained by a joint model of shape and blur. a NRMS fit error of an APC model plotted as a function of NRMS fit error of an APCB model for each neuron. Example cells are filled and labeled. b, d, g Even though the APCB is a generalization of the APC model, i.e., APCB models form a superset of APC models with two additional parameters, cross-validation demonstrates the APCB model to better predict responses to blurred stimuli: b Average NRMSE across training and testing stimuli reveals the APCB model to better predict responses without overfitting; relative error, computed as APC—APCB error, is positive in all significant cases (g). d Relative testing NRMSE between models, plotted as a function of blur selectivity PV, demonstrates the APCB model to better predict responses over a range of neuron tuning profiles. g Significant (black) and not significant (gray) cases of prediction improvement were determined with a pair-wise t-test across hold-out validation stimuli (see Methods: ‘Analysis and model fitting’). c, e, h Observed responses (open circles, dashed lines) and APCB model fits (filled circles, solid lines) for three example neurons (Fig. 2a, c, d). Qualitative assessment of fits suggests that the APCB model captures blur-tuned response properties remarkably well. f Comparison of the APCB gain model against an APCB additive variant with equal degrees of freedom (see Methods: ‘Analysis and model fitting’). Most neurons are better fit by the APCB gain model, particularly when error is small, consistent with blur selectivity being explained by gain modulation
