Confidence resets reveal hierarchical adaptive learning in humans

Area of Science:

• Cognitive Neuroscience
• Computational Neuroscience
• Learning Theory

Background:

• Hierarchical processing is fundamental to brain function, yet its role in learning under uncertainty is debated.
• Both hierarchical and non-hierarchical (flat) models offer explanations for learning, with flat models showing efficiency in dynamic environments.
• Existing methods for differentiating these models based on simple tasks and reported outcomes are insufficient.

Purpose of the Study:

• To critically evaluate the computational significance of hierarchical processing in learning under uncertainty.
• To develop a novel experimental paradigm capable of distinguishing between hierarchical and flat learning models.
• To investigate the utility of confidence reports as a metric for arbitrating between competing learning theories.

Main Methods:

• Development of a complex task with an inherent hierarchical structure enabling generalization across parallel statistics.
• Utilizing confidence reports as a quantitative and qualitative measure to differentiate between hierarchical and flat learning mechanisms.
• Comparison of model predictions against empirical data from the novel task.

Main Results:

• Previously identified hallmarks of hierarchical learning are insufficient for definitive model discrimination.
• The novel complex task successfully elicits behaviors that differentiate between hierarchical and flat learning.
• Confidence reports provide a robust metric for distinguishing between the two learning frameworks.

Conclusions:

• The findings provide strong support for the hierarchical learning framework in cognitive processes.
• Confidence reports serve as a valuable and sensitive metric in the study of learning theory.
• This research offers a more rigorous method for investigating neural processing hierarchies.