Fig. 1: The modeling procedure and the CSVA model.

a BOLD data collected while subjects viewed 1440 images were used for model estimation. Ridge regression was adopted to fit each model to the BOLD time-series for each voxel, using a finite impulse response function with four 2 s time-bins. Weights were estimated for each model feature for each time-bin. These weights characterize each voxel’s response profile or ‘tuning’ to model features. Model validation was conducted using independent fMRI data collected while subjects viewed 180 novel images. Voxel-wise feature weights were used to generate a predicted time-series for each voxel. This was correlated with the recorded BOLD time-series to obtain a metric of model fit that controls for over-fitting, see Methods for details. b The Combined Semantic, Valence and Arousal (CSVA) model comprises 126 mutually exclusive features denoting image semantic category (21 categories), valence (3 levels) and arousal (2 levels) and 18 additional semantic-emotion (SE) compound features that carry both semantic and affective information (e.g. mutilated human; rotten food), see Methods for further details. Image semantic category and SE features were labeled by four independent raters; image valence and arousal were assessed by each subject, see Methods for further details. Here, five example images are labeled with CSVA features. Due to copyright reasons, the images from our database have been replaced with similar images where the photographer, and subject when relevant, have provided consent for the image to be used and publically shared. See table S4 for source & licensing details.