A neural active inference model of perceptual-motor learning

Area of Science:

• Computational Neuroscience
• Cognitive Science
• Robotics

Background:

• The active inference framework (AIF) offers a neuroscientifically grounded approach to understanding human behavior.
• Reward-based learning within AIF can generate human-like actions.
• Anticipation plays a key role in visually guided actions, particularly in tasks like intercepting moving targets.

Purpose of the Study:

• To evaluate the active inference framework's capacity to model anticipatory behavior in human visual-motor tasks.
• To investigate how AIF agents, using artificial neural networks, can replicate human anticipation in target interception.
• To explore the conditions under which anticipatory strategies emerge within the AIF model.

Main Methods:

• A "neural" active inference framework agent was developed using artificial neural networks.
• The agent's actions were selected based on short-term information gain and long-term expected free energy.
• A visual-motor task involving target interception was systematically investigated.
• A novel prior mapping function was formulated to map world-states to free-energy/reward distributions.

Main Results:

• Anticipatory behavior in the AIF agent emerged specifically when movement capabilities were limited.
• Sufficiently long-term estimation of cumulative free energy was crucial for the emergence of anticipatory strategies.
• The proposed AIF model successfully captured anticipatory changes in speed observed in human participants.

Conclusions:

• The active inference framework provides a plausible computational model for anticipatory visually guided behavior in humans.
• Movement constraints and long-term predictive capabilities are key factors influencing anticipatory actions within the AIF.
• The study validates AIF's potential in explaining complex human motor control and decision-making.