Supporting generalization in non-human primate behavior by tapping into structural knowledge: Examples from...

Area of Science:

• Systems Neuroscience
• Computational Neuroscience
• Cognitive Science

Background:

• Complex natural behaviors involve closed-loop action and perception, unlike simplified laboratory tasks.
• Reinforcement learning and control offer a framework to bridge action and perception in neuroscience.

Purpose of the Study:

• To apply reinforcement learning and control to study natural behaviors in neuroscience.
• To guide experimental design by emphasizing active sensing, dynamical planning, and structural regularities.
• To investigate neural circuits underlying flexible and generalizable intelligence.

Main Methods:

• Trained animals to navigate a virtual environment using a joystick, simulating naturalistic behavior.
• Utilized a well-defined, repetitive structure governed by physics to tap into animals' structural knowledge.
• Employed reinforcement learning principles to model decision-making and learning.

Main Results:

• Animals demonstrated zero- or one-shot learning of novel sensorimotor contingencies without further training.
• Inferred evolving latent variables and made decisions consistent with maximizing reward rate.
• Showcased flexible, generalizable, and controlled behaviors for studying intelligence.

Conclusions:

• Reinforcement learning and control are crucial for understanding natural behaviors and guiding neuroscience research.
• Task designs leveraging structural knowledge facilitate studying complex cognitive functions without over-training.
• This approach allows for investigating the neural basis of flexibility, prediction, and generalization.