fMRI data and preprocessing

We used the tfMRI minimally processed data from the Human Connectome Project. For each task, we used the 100 subjects with the lowest mean frame wise displacement. We used AFNI to preprocess the images, matching traditional resting-state functional connectivity studies. The AFNI command 3dTproject was used, passing the mean signal from the cerebral spinal fluid mask, the mean signal from the white matter mask, the mean whole brain signal, and the motion parameters to the “-ort” options, which remove the signals via linear regression. The options “-automask”, which generates the mask automatically was used. The “-passband 0.009 0.08” option, which removes frequencies outside 0.009 and 0.08, was used. Finally, the “-blur 6”, was used, which smooths the images (inside the mask only) with a filter that has a width (FWHM) of 6mm after the time series filtering. Because of the short length of the Emotion task, it was not included in our analyses. Given limitations of the author-topic model in handling negative values, all values below zero were set to zero, allowing us to convert brain activity into "word count".

Fitting the model to the data

To analyze how well the joint probabilistic reverse inference was able to accurately model the original data, we reconstructed the original data using the model estimates of each component's engagement. For each component, we multiplied the component's probability across voxels by the reverse inference's probability for that component being engaged. For example, if the joint probabilistic reverse inference estimated that component 1 has a .1 probability, and particular voxel has a probability of .01 for that component, that voxel is scored as 0.001. We do this for each voxel, for each component, giving us 12 whole brain maps. We then take the sum across the 12 maps (Figure 2). We then calculate the spatial correlation between the original data and this map. We call this value the "fit". To measure if the fit of the model to the data was better than random, we compared the spatial correlation value of the fit of the model to the data to the spatial correlation value of the fit of the model to all other frames (i.e., the data the model was not fit to).