Using serial dependence to predict confidence across observers and cognitive domains

Area of Science:

• Cognitive Psychology
• Neuroscience
• Machine Learning

Background:

• The human perceptual system exploits environmental regularities, leading to serial dependence where past perceptions bias current ones.
• Serial dependence extends to abstract representations like confidence judgments.
• Understanding if confidence patterns generalize across tasks and individuals is crucial for cognitive theories.

Purpose of the Study:

• To investigate the generalizability of temporal patterns in confidence judgments across observers and diverse cognitive domains.
• To determine the key factors influencing serial dependence in confidence generation.
• To explore the relationship between confidence serial dependence and metacognition.

Main Methods:

• Reanalyzed data from the Confidence Database, encompassing perceptual, memory, and cognitive paradigms.
• Employed machine learning classifiers to predict current trial confidence based on historical confidence data.
• Utilized cross-observer and cross-domain decoding to assess model generalizability.

Main Results:

• A machine learning model trained on perceptual confidence data successfully predicted confidence across different observers and cognitive domains.
• The recent history of confidence judgments was the most significant predictor of current confidence.
• Confidence predictions generalized across both correct and incorrect trials, indicating a dissociation from metacognitive evaluation.

Conclusions:

• Temporal patterns in confidence judgments exhibit domain-general and cross-observer serial dependence.
• Serial dependence in confidence is primarily driven by recent confidence history, independent of accuracy or reaction time.
• Confidence generation's serial dependence appears uncoupled from metacognition, contributing to the debate on domain-specificity versus domain-generality in cognitive processes.