Experience-driven recalibration of learning from surprising events

Area of Science:

• Cognitive Science
• Neuroscience
• Machine Learning

Background:

• Adaptive learning is crucial for navigating dynamic environments.
• The influence of surprising events on learning can vary significantly across contexts.
• Understanding how individuals regulate learning based on surprise is key to explaining flexible behavior.

Purpose of the Study:

• To investigate whether humans spontaneously regulate the impact of surprise on their learning.
• To examine how individuals adapt their learning rates in response to surprising events in different environments.
• To determine if participants can recalibrate their learning strategies when environmental statistics change.

Main Methods:

• Two experiments involving a perceptual judgment task under spatial uncertainty.
• Analysis of uninstructed predictive eye movements to assess learning patterns.
• Observing participants' responses to environments with varying statistical properties and unsignaled reversals.

Main Results:

• Participants demonstrated a form of metalearning, adjusting learning rates based on the informativeness of surprising events.
• Predictive eye movements modulated learning rates in opposite directions across different contexts.
• Participants successfully readjusted learning patterns when environmental statistics were reversed, though performance was better when surprises indicated meaningful change.

Conclusions:

• Humans exhibit spontaneous, context-appropriate recalibration of surprise's role in adaptive learning.
• There may be a bias towards interpreting surprise as informative or difficulty ignoring salient outliers.
• These findings highlight the flexibility of human learning in dynamic environments.