Learning to select computations in recurrent neural circuits

Area of Science:

• Computational neuroscience
• Cognitive science
• Artificial intelligence

Background:

• Biological computation exhibits flexibility and efficiency, often linked to cognitive control.
• The neural mechanisms underlying adaptive control balancing utility and computational cost remain largely unknown.
• Prefrontal cortex is implicated in higher-level cognitive functions and decision-making.

Purpose of the Study:

• To propose a computational framework for adaptive control of thought.
• To model how the brain implements flexible and efficient cognitive processes.
• To unify theories of meta-reasoning and meta-learning in neural systems.

Main Methods:

• Developed a recurrent neural network model integrating rational meta-reasoning theory and a meta-learning algorithm.
• Utilized the model to simulate performance in simple choice and multi-step planning tasks.
• Compared model outputs with neural dynamics from macaque orbitofrontal cortex and human planning behaviors.

Main Results:

• The model successfully learned to select computations, approximating optimal symbolic models in choice tasks.
• It reproduced neural dynamics observed in macaque orbitofrontal cortex.
• In planning tasks, the model replicated human behavioral strategies and associated neural dynamics.

Conclusions:

• The proposed framework offers a mechanistic account of adaptive control of thought.
• Learning to reason can be framed as learning to learn from internal cognitive operations.
• This work bridges computational theory and neural implementation of cognitive control.