A number simplex in the human medial temporal lobe.

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Neuroscience

Background:

• The prevalent mental number line model may not fully capture the complexity of human numerical cognition.
• Neural representations of numbers are crucial for understanding mathematical abilities.

Purpose of the Study:

• To investigate the neural manifold structure underlying numerical cognition in humans.
• To compare human neural representations with those in large language models (LLMs).
• To explore the neural basis of arithmetic processing and its relation to mathematical capacity.

Main Methods:

• Recording neural activity from medial temporal lobe (MTL) neurons in humans during dot counting and arithmetic tasks.
• Analyzing population codes to identify manifold geometry.
• Comparing human neural data with representations in LLMs.
• Decoding arithmetic results from neural activity.

Main Results:

• Neural coding of numerosity in the MTL forms high-dimensional, simplex-shaped manifolds.
• These simplicial manifolds offer greater flexibility and expressivity compared to linear models.
• Distinct simplicial population codes were observed for dot arrays and Arabic numerals.
• Similar simplicial geometry was found in LLMs, suggesting convergent representational strategies.
• Arithmetic results were decodable from neural activity, with accuracy correlating with mathematical capacity.
• Linear transformations of simplicial representations modeled arithmetic processing, akin to LLM attention mechanisms.

Conclusions:

• Human numerical cognition is supported by high-dimensional, simplex-shaped neural manifolds.
• This representational geometry provides a flexible foundation for complex numerical tasks.
• The findings suggest potential parallels between human arithmetic processing and LLM architectures.