Spontaneous representation of numerosity zero in a deep neural network for visual object recognition

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Neuroscience

Background:

• The concept of zero as a numerical value is a complex cognitive ability shared by humans and some advanced animals.
• Understanding the origins of numerical representation, particularly for zero, is crucial for both AI and cognitive science.

Purpose of the Study:

• To investigate whether a deep learning neural network can spontaneously develop a representation of zero without explicit numerical training.
• To identify neural network signatures analogous to biological neurons that indicate the processing of numerical quantity, specifically zero.

Main Methods:

• A deep learning neural network was trained on visual information without any numerical data.
• The network's internal representations were analyzed to identify units responsive to empty sets.
• Specific network units were deactivated to assess their functional role in zero representation.

Main Results:

• The neural network spontaneously developed units that represent zero, showing maximum activity for empty sets and decreasing activity with increasing numerosities.
• These findings suggest that the network ordered zero as the smallest numerical value.
• Selective removal of these units led to deficits in the network's zero-judgment capabilities, confirming their functional relevance.

Conclusions:

• Visual information processing alone is sufficient for a sense of number, including the representation of zero, to emerge in artificial neural networks.
• This emergent property in AI may explain the rudimentary understanding of zero observed in cognitively advanced animals with shared nonverbal number systems.