Computer Vision News - July 2020

2 Summary Research 6 To understand the function of this network, three patches of total 1244 images, made of 51 objects belonging to 6 different categories were recorded. The results showed that all the responses were very consistent between those categories. A common general anatomical organization is observed with coding scheme similarity between the body, the NML and the face networks. The stimuli showed to the monkey were project onto the first two dimensions of the object space while marking the top 100 images for the three networks. This confirmed the hypothesis of valid topological organization of the patches. The activations of units found in the 8 layers of the deep network. With thousands of units in each layer, it wasn’t straightforward to discern any patterns to their firing. As they spanned across three quadrants of the space and one of the principal components was strongly activated by spiky objects, a prediction that a fourth network exists was consequently made (for stubby, inanimate objects see Fig. 1). Figure 1: The four quadrants of object space derived from a deep neural network are shown. These are the four quadrants as described in the paper (inanimate, stubby, animate, spiky) represented by four distinct networks in the primate brain.