Learned predictiveness models predict opposite attention biases in the inverse base-rate effect.

Area of Science:

• Cognitive Psychology
• Computational Neuroscience
• Machine Learning

Background:

• Attention-based models of associative learning often rely on the learned predictiveness principle.
• These models aim to optimize learning by focusing on the most predictive features.
• Existing models differ in their mechanisms, leading to varied predictions, particularly for the inverse base-rate effect.

Purpose of the Study:

• To investigate how attention-based models account for the inverse base-rate effect in associative learning.
• To compare the predictive accuracy of the modified Mackintosh and EXIT models using simulations and human data.
• To examine attentional dynamics during causal learning, including the influence of context and feedback.

Main Methods:

• Simulations were conducted using the modified Mackintosh and EXIT models.
• A human causal learning task was designed to replicate the inverse base-rate effect.
• Eye-tracking was employed to measure attention shifts during decision-making and feedback phases.

Main Results:

• Both models could explain rare-outcome choice biases in the inverse base-rate effect.
• The EXIT model accurately predicted that rare predictors receive greater attention than common predictors throughout training.
• Attentional patterns differed between pre-decision and feedback phases, with context effects observed only pre-decision.

Conclusions:

• The EXIT model provides a better account of human attentional dynamics in causal learning than the modified Mackintosh model.
• Learned predictiveness is crucial, but specific model mechanisms dictate predictions for attention and learning.
• Contextual information influences attention differently depending on whether it precedes a decision or follows feedback.