Representational invariance of numerosity via divisive normalization in coarse and fine visual channels

Robust evidence now suggests that numerosity perception emerges from the early visual cortex. However, such empirical findings pose a theoretical challenge for explaining how a low-level perceptual system represents discrete values from continuous input independently of other magnitude dimensions. Among proposals for this representational invariance, two computational accounts with ties to neural data have garnered attention: divisive normalization and Fourier decomposition. Here, we test these hypotheses using an integrated neural, behavioral, and computational approach and show that: (1) The visual cortex remains sensitive to numerosity even when the input images are equalized for their Fourier power, inconsistent with the Fourier decomposition account. (2) The divisive normalization model explains this neural phenomenon through the selective disruption of fine but not coarse visual channels when encoding normalized local contrast. (3) Backward masking that disrupts fine processing in a psychophysical experiment degrades the acuity of intact dot arrays to the level of Fourier-power equalized dot arrays, which validates the unique prediction of the divisive normalization model. These findings provide converging evidence for the proposal that normalized local contrast enables representational invariance of numerosity in the visual cortex.