Demystifying excessively volatile human learning: A Bayesian persistent prior and a neural approximation.

Area of Science:

• Neuroscience
• Cognitive Psychology
• Computational Modeling

Background:

• Humans and animals learn statistical regularities for prediction, a key problem in neuroscience and psychology.
• The Dynamic Belief Model (DBM) suggests humans default to assuming abrupt environmental changes, even in stable conditions.
• Exact Bayesian inference for switching state-space models is computationally intensive, leading to research in approximate algorithms.

Purpose of the Study:

• To examine the neurally plausible exponential filtering (EXP) algorithm for approximating Bayesian predictions.
• To compare EXP's performance against other algorithms and human behavior.
• To investigate the role of persistent prior influence in human learning and prediction.

Main Methods:

• Derived the theoretical relationship between the Dynamic Belief Model (DBM) and exponential filtering (EXP).
• Evaluated EXP's computational efficiency and its ability to approximate Bayesian prediction performance.
• Compared model performance against human behavior in a visual search task.

Main Results:

• EXP is a computationally efficient algorithm, simpler than most alternatives except the delta-learning rule.
• EXP significantly outperforms the delta-learning rule in approximating Bayesian prediction performance.
• EXP reproduces human behavior in a visual search task comparably to DBM, suggesting a persistent prior influence.

Conclusions:

• The exponential filtering (EXP) algorithm offers a computationally efficient yet near-Bayesian approach to prediction.
• Human learning and prediction likely incorporate a persistent prior influence, as evidenced by EXP's success.
• In information-poor environments, detecting environmental changes or modulating learning rates is often unnecessary for optimal predictions.