Unifying the structures of language in a neural population code

Area of Science:

• Cognitive Neuroscience
• Artificial Intelligence
• Computational Linguistics

Background:

• Traditional rule-based systems struggle with natural language's complexity.
• Large language models (LLMs) utilize statistical learning for language processing.
• LLMs encode linguistic structures in high-dimensional embedding spaces.

Purpose of the Study:

• To explore LLMs as cognitively plausible models of human language processing.
• To investigate the parallels between LLM representations and neural population codes.
• To propose new theoretical frameworks for understanding language emergence.

Main Methods:

• Analyzing the computational strategies of LLMs.
• Comparing LLM representations to neural population codes in cognitive neuroscience.
• Synthesizing research at the intersection of AI and neuroscience.

Main Results:

• LLMs exhibit a context-driven, distributed representation of language.
• This representation closely parallels neural population codes observed in the brain.
• LLMs offer a viable computational model for neural language processing.

Conclusions:

• LLMs provide insights into the neural computations underlying human language.
• Understanding language emergence requires new theoretical paradigms beyond symbolic systems.
• Similar computational strategies may underlie language in both biological and artificial systems.