Structure learning enhances concept formation in synthetic Active Inference agents

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Artificial Intelligence

Background:

• Humans excel at environmental learning, forming flexible concepts from context-specific affordances and relations.
• Understanding concept formation is key to artificial intelligence and cognitive modeling.

Purpose of the Study:

• To present a deep hierarchical Active Inference model for goal-directed behavior and concept learning.
• To elucidate mechanisms underlying concept learning in spatial foraging tasks using simulations.

Main Methods:

• Developed a deep hierarchical Active Inference model.
• Employed simulations of a spatial foraging task.
• Investigated belief update schemes via maximizing model evidence.

Main Results:

• Model representations reflect environmental structure, enhanced by Bayesian model reduction.
• Synthetic agents learned associations and formed concepts via inferential, parametric, and structure learning.
• Learned representations captured environmental symmetries.

Conclusions:

• Active Inference provides a framework for understanding concept formation and environmental abstraction.
• Bayesian model reduction plays a crucial role in refining learned representations.
• The model demonstrates how diverse beliefs and structures emerge from learning processes.