Transitive inference as probabilistic preference learning.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Machine Learning

Background:

• Transitive inference (TI) involves inferring new relationships from known ones.
• TI exhibits behavioral signatures like the serial position effect (SPE) and symbolic distance effect (SDE).
• The brain's ability to manage and integrate ranking models is crucial for TI.

Purpose of the Study:

• To propose a novel framework for understanding transitive inference (TI).
• To model TI as a probabilistic preference learning task using Mallows models.
• To explore the neural underpinnings of TI through computational modeling.

Main Methods:

• Utilized one-parameter Mallows models to represent TI as a probabilistic preference learning task.
• Conducted simulations to validate the Mallows model's effectiveness.
• Extended the model with Bayesian selection for hypothesis generation and merging.
• Employed neural networks to replicate Mallows models and compare with neural data.

Main Results:

• The Mallows ranking model successfully reproduced the symbolic distance effect (SDE) and serial position effect (SPE).
• Bayesian extension demonstrated the model's capability to generate and merge ranking hypotheses.
• Neural network replication showed alignment with prefrontal neural activity during TI.

Conclusions:

• The proposed probabilistic preference learning framework offers a new perspective on transitive inference (TI).
• Mallows models provide a robust computational tool for explaining TI phenomena.
• This approach bridges computational modeling and neuroscience to elucidate TI mechanisms.