Visual number sense in untrained deep neural networks

Area of Science:

• Neuroscience
• Artificial Intelligence
• Cognitive Science

Background:

• Number sense, or numerosity estimation, is present in animals but its neural basis is not well understood.
• The emergence of abstract number sense, independent of visual cues, is a key question in cognitive neuroscience.

Purpose of the Study:

• To investigate the spontaneous emergence of number-selective neurons in a deep neural network modeling the ventral visual stream.
• To determine if these neurons can account for abstract number sense and number comparison abilities.

Main Methods:

• Utilized a deep neural network simulating the brain's ventral visual stream.
• Analyzed neuronal activity in a randomly initialized network without explicit learning.
• Examined network performance on number comparison tasks.

Main Results:

• Number-selective neurons emerged spontaneously in the randomly initialized network.
• Neuronal activity patterns, combining increasing and decreasing functions, drove number tuning.
• The network demonstrated abstract number sense and replicated neural characteristics observed in biological brains.

Conclusions:

• Innate cognitive functions like number sense may arise spontaneously from the statistical properties of neural network connectivity.
• This model provides a potential explanation for the origin of number sense, independent of explicit learning.
• The findings offer insights into the neural underpinnings of innate cognitive abilities.