Trained deep neural network models of the ventral visual pathway encode numerosity with robustness to object and scene identity

“Number sense”, the ability to quickly estimate quantities of objects in a visual scene, is present in humans and many other animals, and has recently been demonstrated in biologically inspired vision models, even before training. However, real-world number perception requires abstraction from the properties of individual objects and their contexts, in contrast to the simplified dot patterns used in previous studies. Using novel, synthetically generated photorealistic stimuli, we discovered that deep convolutional neural networks optimized for object recognition can encode numerical information across varying object and scene identities in their distributed activity patterns. In contrast, untrained networks failed to discriminate numbers, and appeared to encode low-level visual summary statistics of scenes rather than the number of discrete objects per se. These results caution against using untrained networks to model early numerical abilities and highlight the need to use more complex stimuli to understand the mechanisms behind the brain’s visual number sense.