Fig. 9: OTC tuning patterns predict behavioral responses to emotional natural stimuli.

a We examined the extent to which OTC tuning to emotional natural images, as captured by CSVA group-level PC scores, predicted the behavioral responses selected for each image by MTurk raters, across images. The dotted red line represents the percentage of out-of-sample variance in behavioral responses explained as a proportion of potential explainable variance (y axis) plotted against the number of PCs included as predictors in the ordinary least squares regression analysis. The error band around the dotted line represents the 95% confidence interval (CI). We also calculated the scaled¹ out-of-sample variance in behavioral responses explained using PCs derived directly from PCA on CSVA image features, across images (yellow dotted line; 95% CI is given by the associated error band). Across all levels of dimensionality considered (nu. of PCs=1 to 21), OTC tuning to CSVA features predicted behavioral responses significantly better than components from PCA conducted directly on the features themselves. This is consistent with OTC showing selective representation of image semantic and affective features pertinent to behavior. b Here, data are shown using the same format as in a Dotted lines represent scaled¹ out-of-sample variance in behavioral responses explained by PCs obtained from PCA on OTC feature weights for the CSVA model (red) versus a Gabor model (pink), the Semantic Only model (dark blue) and the Valence by Arousal model (light blue). Error bands give the 95% CIs. Given the smaller feature space of the Valence by Arousal model, the maximum number of PCs that can be extracted for this model is six. The poor performance of Gabor model PCs in predicting behavioral responses suggests that OTC tuning to low-level image structural features is insufficient to guide behavior. Both the Semantic Only and Valence by Arousal models outperform the Gabor model in predicting behavior. However, their maximal prediction of behavior (at n = 21 and n = 6 PCs respectively) is significantly less than that achieved by the CSVA model using an equivalent number of components. ¹Scaled by potential explainable variance (see methods).