Intrinsically motivated action-outcome learning and goal-based action recall: a system-level bio-constrained...

Area of Science:

• Computational Neuroscience
• Animal Behavior
• Machine Learning

Background:

• Animal learning is driven by extrinsic motivations (EMs) for survival and reproduction, and intrinsic motivations (IMs) for acquiring actions without immediate rewards.
• While IMs have been studied in psychology and neuroscience, current computational models lack integration of key aspects and neurobiological constraints.
• Existing models in computational modeling, robotics, and machine learning capture only certain facets of IMs.

Purpose of the Study:

• To propose a bio-constrained system-level model that integrates key aspects of intrinsically motivated learning and behavior.
• To elucidate the neural mechanisms underlying intrinsic motivation, focusing on action-outcome association, attention, and reward recall.
• To provide a framework for guiding empirical experiments and computationally validating theories on intrinsic motivations.

Main Methods:

• Developed a system-level computational model incorporating neurobiological constraints.
• Focused on three core processes: acquisition of action-outcome associations via dopamine signals, transient visual/action focusing, and goal-directed recall of actions.
• Tested the model using simulations, including selective lesion experiments, to assess learning speed and goal-directed behavior recruitment.

Main Results:

• The model demonstrates that focusing processes enhance the speed of learning action-outcome associations.
• Validated that these learned associations can be effectively recruited for goal-directed behaviors.
• Selective lesion tests confirmed the model's predictions regarding the role of specific neural mechanisms.

Conclusions:

• The proposed bio-constrained model represents a significant step towards a unified understanding of intrinsic motivation.
• The model provides a computational framework that can guide future research in neuroscience and artificial intelligence.
• Findings highlight the importance of integrating neurobiological constraints into computational models of learning and motivation.