Spontaneous generation of innate number sense in untrained deep neural networks

Abstract

Number-selective neurons are observed in numerically naive animals, but it was not understood how this innate function emerges in the brain. Here, we show that neurons tuned to numbers can arise in random feedforward networks, even in the complete absence of learning. Using a biologically inspired deep neural network, we found that number tuning arises in three cases of networks: one trained to non-numerical natural images, one randomized after trained, and one never trained. Number-tuned neurons showed characteristics that were observed in the brain following the Weber-Fechner law. These neurons suddenly vanished when the feedforward weight variation decreased to a certain level. These results suggest that number tuning can develop from the statistical variation of bottom-up projections in the visual pathway, initializing innate number sense.