Understanding human metacontrol and its pathologies using deep neural networks.

Area of Science:

• Cognitive Neuroscience
• Computational Psychiatry
• Artificial Intelligence

Background:

• Error monitoring is essential for assessing environmental controllability and estimating the value of metacontrol.
• Metacontrol involves higher-order cognitive processes that regulate lower-level control mechanisms.
• Understanding the neural and behavioral correlates of metacontrol is crucial for both cognitive science and clinical applications.

Purpose of the Study:

• To investigate the behavioral and neural correlates of error monitoring and metacontrol using computational simulations.
• To explore the utility of deep neural networks (DNNs) as a model system for studying metacontrol.
• To examine how misestimations of environmental controllability impact behavior and potentially model human psychological traits.

Main Methods:

• Utilized deep reinforcement learning (RL) agents and human participants in a reward-guided learning task.
• The task involved adapting to dynamic changes in action controllability.
• Trained RL agents to explicitly predict action prediction errors, mimicking medial prefrontal cortex activity.

Main Results:

• RL agents successfully performed the task only when trained to predict action prediction errors.
• Trained RL agents exhibited metacontrol signatures comparable to those observed in humans.
• Over- or underestimating controllability in RL agents led to behavioral pathologies mirroring human depressive, anxious, or compulsive traits.

Conclusions:

• Deep neural networks can serve as a valuable tool for modeling metacontrol processes.
• Explicit prediction of action prediction errors is critical for developing metacontrolling AI agents.
• Computational models of metacontrol can offer insights into the mechanisms underlying human psychological conditions.