Figure 3

Visual similarity tree for our trained CNN. The similarity tree is based on hierarchical clustering of the response of the last fully-connected layer in our trained CNN to 6000 randomly selected training images of particular species (i.e., feature vectors of the images). The leaves represent feature vector centroids of 300 training images of each species, and their relative positions in the tree indicate the Euclidean distances between these centroids in the feature space. In the similarity tree, the more similar the response of this layer to two species, the more tightly coupled they are in the tree. Green, purple, and brown branches correspond to three primary clusters that appear to be a small to medium-sized antelope cluster, an animals-with-prominent-tail or big-ears cluster (though baboons seem to be an outlier in this group), and a relatively large body-to-appendages group (with waterbuck the outlier in this group). When the feature vectors of unkown animal species are placed in the tree (e.g., the red branch of lion), sometimes they can differ greatly from those of the known species.