Generative models, linguistic communication and active inference.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Cognitive Science

Background:

• Active inference provides a framework for understanding goal-directed behavior.
• Simulating complex cognitive processes like language requires sophisticated generative models.
• Dyadic interactions, or communication between two agents, present unique modeling challenges.

Purpose of the Study:

• To develop a biologically plausible generative model for simulating communication between synthetic agents.
• To explore generative models capable of supporting dialogue and linguistic exchange.
• To investigate emergent properties of communication within an active inference framework.

Main Methods:

• Utilized active inference formulations to model dyadic interactions.
• Developed deep, hierarchical generative models with high-order interactions among discrete states.
• Simulated linguistic communication using a 'Twenty Questions' game with synthetic subjects.

Main Results:

• Demonstrated that generative models with specific factorial structures can accommodate language's combinatorial nature.
• Showcased emergent behavioral and neurophysiological correlates of linguistic communication through variational message passing.
• Identified theta-gamma coupling as an emergent property of belief updating during listening.

Conclusions:

• The proposed generative model and inference scheme offer a biologically plausible approach to simulating communication.
• The model successfully illustrates key architectural principles for dialogue systems.
• Emergent properties, such as neural oscillations, can arise from fundamental principles of belief updating in communication.