A neural network for modeling human concept formation, understanding and communication

Area of Science:

• Cognitive Neuroscience
• Computational Neuroscience
• Artificial Intelligence

Background:

• The human brain's ability to form abstract concepts from sensory experiences is crucial for flexible cognition but its computational basis is unclear.
• Understanding this process is vital for both cognitive science and developing advanced AI systems.

Purpose of the Study:

• To propose a computational framework, CATS Net, that models the mechanism of abstract conceptual representation and flexible application.
• To investigate how conceptual structures enable knowledge transfer and relate to neural mechanisms in the human brain.

Main Methods:

• Developed a dual-module neural network (CATS Net) with concept-abstraction and task-solving modules.
• Implemented hierarchical gating control by extracted concepts for task performance.
• Performed model-brain fitting analyses comparing CATS Net's representations with human neuroimaging data.

Main Results:

• CATS Net successfully extracted low-dimensional conceptual representations from sensorimotor data.
• The model demonstrated transferable semantic structures enabling cross-network knowledge transfer.
• Emergent concept spaces in CATS Net aligned with human ventral occipitotemporal cortex activity and neurocognitive semantic models.
• Gating mechanisms in the model mirrored human semantic-control brain networks.

Conclusions:

• CATS Net provides a unified computational framework for understanding human conceptual cognition.
• The model offers mechanistic insights into how abstract concepts are formed and utilized.
• This work advances the engineering of artificial systems with human-like conceptual intelligence.