Investigation of Numerosity Representation in Convolution Neural Networks

Convolutional neural networks (CNNs) have emerged as powerful models for predicting neural activity and behavior in visual tasks. Recent studies suggest that number-detector units—analogous to number neurons—can emerge in CNNs, both in trained networks optimized for object recognition and in untrained networks. In this work, we extend previous studies by investigating whether CNNs encode numerosity at the population level and by examining how the statistical distribution of numerical and non-numerical features in the training dataset influences their internal representations. Recognizing that perceptual systems are finely tuned to the statistical properties of their sensory environment, we compare CNNs trained on both synthetic datasets and a naturalistic dataset that better reflects the real-world conditions shaping human number sense. By systematically manipulating these statistical properties, we assess their impact on the encoding of both numerical and non-numerical features. Finally, we compare these computational representations with those observed in the human brain, highlighting both shared characteristics and key differences that provide deeper insights into the mechanisms underlying numerosity perception in biological and artificial systems.