Figure 2.

(A) Topic probabilities for the Wikipedia articles about the 60 concepts for which we have fMRI data. Each concept belongs to one of 12 semantic categories, and concepts are grouped by category (five animals, five insects, etc.). Note that the category structure visible is due to how we sorted the columns for display; the model is trained in an unsupervised manner and knows nothing about category structure. Note also that there are topics that are not probable for any of the concepts, which happens because they are used for other concepts in the 3500 concept corpus. Below this are the top 10 most probable words in the probability distributions associated with three of the topics. (B) The decomposition of the brain image for “House” into a weighted combination of topic basis images. The weights allow us to combine the corresponding topic word distributions into an overall word distribution (top 10 words shown).