Fig. 2: Relating visual and semantic features.
From: Computational reconstruction of mental representations using human behavior
a The 100 most common responses provided by participants. The size of the word is approximately proportional to its frequency. b Illustration of the significant associations between the first 60 CNN principal components (PCs) and the first 30 semantic PCs (p < 0.05, randomization test, two-tailed, FWER-corrected). c Detail of the significant associations between the first 6 semantic PCs and the first 8 CNN PCs. The color of the curves represents the strength and polarity of the association (see scale in (b)), and the thickness is proportional to its absolute value. Semantic PCs are summarized by the 5 words provided in the experiment that loaded most strongly (word size approximately proportional to closeness). CNN PCs are summarized by both a synthesized image maximizing the value of that PC (top) and the 4 stimuli that loaded most strongly (bottom). d Summary of the semantic content associated with the top 8 CNN PCs. The 5 words that corresponded most closely to the semantic feature values associated with each CNN PC are shown (word size approximately proportional to closeness).
