Number, texture and crowding.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Computer Vision

Background:

• The 'sense of number' or numerosity is a key cognitive ability.
• Its neural and computational underpinnings, particularly in visual perception, remain incompletely understood.
• Existing models often link numerosity to texture or spatial properties, necessitating further investigation into its independence.

Purpose of the Study:

• To investigate whether a 'sense of number' can emerge intrinsically within artificial neural networks designed for visual perception.
• To determine if numerosity is an emergent property of hierarchical generative models.
• To explore the independence of numerosity from texture-based visual attributes.

Main Methods:

• Utilized deep networks based on hierarchical generative models of visual perception.
• Analyzed the network's internal representations and emergent properties during visual processing tasks.
• Compared network selectivity for numerosity with its processing of texture and crowding phenomena.

Main Results:

• Selectivity for numerosity was observed as a natural emergent property in the deep network models.
• The findings indicate that the network developed a 'sense of number' without explicit training for this task.
• Numerosity processing appeared independent of texture-related features and other potentially confounding visual attributes.

Conclusions:

• Hierarchical generative models of visual perception can naturally develop selectivity for numerosity.
• This suggests that the 'sense of number' may be a fundamental visual property, not solely reliant on texture or related attributes.
• The study provides a computational framework for understanding the emergence of numerical cognition within visual systems.