On Numerosity of Deep Neural Networks

Xi Zhang, Xiaolin Wu

Introduction

In cognitive science, "numerosity" refers to the innate ability to perceive the quantity of objects in a set without explicit counting. This fundamental skill is observed in human infants and various animal species. Recent research sparked significant interest by claiming that this "number sense" spontaneously emerges in Deep Convolutional Neural Networks (DCNNs) trained solely for visual object recognition on datasets like ImageNet. This finding suggested that abstract mathematical concepts might be a natural byproduct of learning to perceive the visual world. However, a critical re-examination by Zhang and Wu challenges this notion, suggesting that the observed "number neurons" may be statistical artifacts rather than emergent properties of intelligence.

Methodology and Datasets

To investigate these claims, the researchers employed a two-pronged approach. First, they re-evaluated the statistical methods used to identify numerosity-selective neurons. Second, they trained a dedicated network, "Nu-Net," specifically for numerosity tasks to test its ability to generalize abstract concepts.

The study utilized binary images of dot patterns across three stimulus sets designed to control for low-level visual features that might correlate with quantity:

  • Standard Set: Circular dots with random sizes and spacing.
  • Control Set 1: Dots adjusted to have equal total area and density across different quantities.
  • Control Set 2: Various geometric shapes with equal "convex hull" areas.
Figure 1: Examples of stimulus sets used to control for low-level visual features across different numerosities.

The Statistical Artifact: The "Small Sample" Trap

The core of the critique lies in the statistical tool used to identify "number neurons": the Analysis of Variance (ANOVA). The original study used a relatively small sample of 336 test images to evaluate over 37,000 neurons. Zhang and Wu demonstrate that with such a massive search space and a limited sample size, it is statistically inevitable to find neurons that appear to respond to specific numbers purely by chance.

When the researchers increased the ANOVA sample size, the results changed dramatically. In the object recognition network, the percentage of identified "number neurons" plummeted from approximately 10% to zero as the sample size per combination grew from 7 to 80 images.

Figure 3: Percentage of numerosity-selective units as a function of ANOVA sample size for (a) an object recognition network and (b) a randomly initialized network.

Crucially, as shown in Figure 3, a randomly initialized (untrained) network actually exhibited a higher percentage of "number-selective" neurons (12.8%) than the trained network (9.6%) at small sample sizes. This finding strongly suggests that these neurons are not a product of learning but are instead statistical noise.

Generalization and the Subitizing Effect

Beyond the statistical critique, the authors tested whether a network trained specifically for numerosity (Nu-Net) could truly grasp the abstract concept of number. While Nu-Net achieved high accuracy (95.6%) on data similar to its training set, its performance collapsed when tested on "out-of-distribution" images. When object sizes were varied by 50%, accuracy dropped to 64.2%. This indicates that the network was likely learning specific visual patterns (e.g., total pixel density) rather than an abstract mathematical concept.

However, the researchers noted a "subitizing" effect: the networks remained remarkably robust when identifying very small quantities (1 to 4). For these small numbers, the network maintained a perfect estimation interval, mirroring the human ability to instantly perceive small sets without counting. For larger quantities (e.g., 24-30), uncertainty increased significantly.

Conclusion

The study by Zhang and Wu provides a necessary reality check on the "spontaneous emergence" of number sense in DCNNs. By demonstrating that previous findings were likely artifacts of small sample sizes and that even dedicated networks struggle to generalize, the authors highlight the gap between pattern matching and true abstract reasoning in current AI architectures. While the robustness of small-number recognition (subitizing) offers an interesting parallel to biological systems, achieving true abstract numerosity in artificial systems remains an open challenge for future research.

Summary updated

BibTeX
@inproceedings{zhang2020numerosity,
  author = {Xi Zhang and Xiaolin Wu},
  title = {{On Numerosity of Deep Neural Networks}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2020},
  url = {https://openreview.net/forum?id=telbpEar6b},
}
Figure