Fig. 5: Analysis of class labels for synthesized images. | Nature Communications

Fig. 5: Analysis of class labels for synthesized images.

From: Direct perception of affective valence from vision

Fig. 5

a Distribution of the average neural objective function for the 1000 Alexnet classes. The top 50 largest and top 50 smallest classes represent entities evoking the most similar visual activity to the VV+ and VV− masks, respectively. b Hierarchical tree-structure semantic analysis of the labels of the top 50 positive (orange) and top 50 negative classes (blue). Horizontal gray bars represent the proportion of classes across the ImageNet’s 1000 class list. Nodes are expanded as long as they correspond to at least 5% of the list (>20 classes). The number of classes that fall into each node of the tree is shown on arrows (blue with ‘−’ showing the count out of the lowest 50, and orange with ‘+’ showing the count out of the highest 50 classes). ‘Artifact’ versus ‘Living thing’ are the main nodes separating the two label sets. Source data are provided as a Source Data file.

Back to article page