Number detectors spontaneously emerge in a deep neural network designed for visual object recognition

Area of Science:

• Neuroscience
• Computer Science
• Cognitive Science

Background:

• Humans and animals possess an innate
• number sense,
• enabling intuitive quantity assessment.
• This suggests that the brain's visual system, crucial for object recognition, inherently processes numerosity.

Purpose of the Study:

• To investigate if number sense mechanisms can spontaneously emerge within artificial neural networks.
• To determine if a deep neural network trained solely on visual object recognition develops units sensitive to abstract numerosity.

Main Methods:

• A biologically inspired deep neural network was trained on visual object recognition tasks.
• The internal network units were analyzed for tuning to abstract numerosity.
• The network's number discrimination performance was evaluated against established psychophysical laws.

Main Results:

• Spontaneously emerging network units exhibited tuning to abstract numerosity, akin to biological number neurons.
• The network demonstrated number discrimination performance consistent with the Weber-Fechner law.
• These findings indicate that numerosity processing is an emergent property of visual recognition mechanisms.

Conclusions:

• The number sense can spontaneously emerge from neural network architectures trained on visual recognition.
• The visual system's inherent mechanisms are sufficient for developing number sense capabilities.
• This research provides a computational model for understanding the biological basis of number sense.