Fig. 3. Single unit properties of deep networks.

a Schematic illustrating the general principle observed in neurons in high-level visual areas of the brain. The response of a neuron to multiple objects is typically the average of its responses to the individual objects at those locations. b Response to multiple-object displays plotted against the sum of the individual object responses for two-object displays (black) and three-object displays (red), across 10,000 units randomly selected from Layer 37 of the VGG-16 network. c Normalization slope plotted across layers for two object displays (blue) and three-object displays (brown) for the VGG-16 network and a randomly initialized VGG-16 (brown). The dashed lines depict the slopes observed in monkey IT neurons using a different stimulus set²¹. d Selected stimuli used to compare sparseness across multiple dimensions in a previous study of IT neurons²². e Correlation between sparseness on the reference set vs. along morphlines across units of each layer in the VGG-16 network (blue) and a VGG-16 with random weights (brown). The dashed line indicates the observed correlation in monkey IT neurons on the same set of stimuli. f Correlation between sparseness for textures and sparseness for shapes plotted across layers of the VGG-16 network (blue) and a VGG-16 with random weights (brown). The dashed line indicates the observed correlation in monkey IT neurons on the same set of stimuli.