Intermittent Active Inference

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Cognitive Science

Background:

• Active inference models perception and action via prediction error minimization.
• Human control strategies are often intermittent, reducing computational load and noise.
• Standard active inference assumes continuous processing, which may be computationally demanding.

Purpose of the Study:

• Introduce Intermittent Active Inference (IAIF) to model intermittent control strategies.
• Investigate the efficacy of intermittent planning within the IAIF framework.
• Evaluate IAIF's performance and efficiency against continuous planning.

Main Methods:

• Developed IAIF where sensing, inference, planning, or acting can occur intermittently.
• Focused on intermittent planning: agents re-plan only when prediction error exceeds a threshold or Expected Free Energy (EFE) surpasses estimates.
• Evaluated IAIF in a mouse pointing task, comparing against continuous planning.

Main Results:

• IAIF significantly reduces computation time compared to continuous planning.
• Task performance is maintained or improved with IAIF, especially when increasing sampled plans.
• The EFE-based trigger for re-planning requires no additional calibration.

Conclusions:

• Intermittent planning within IAIF offers a computationally efficient approach to active inference.
• IAIF effectively models intermittent control, reducing computational demands while preserving performance.
• IAIF is a valuable and practical extension for computational modeling workflows.