Modeling relational responding with artificial neural networks

Area of Science:

• Comparative psychology
• Artificial intelligence
• Cognitive science

Background:

• Relational responding involves behavior aligned with rules for comparing stimuli.
• Previous studies trained pigeons on stimulus comparison tasks with mixed results.
• Contrasting theories exist regarding the mechanisms of relational responding in animals.

Purpose of the Study:

• To develop and validate an artificial neural network model of relational responding.
• To analyze the model's generalization capabilities from training to test stimuli.
• To reconcile theoretical perspectives on relational responding.

Main Methods:

• Trained pigeons on a task involving choosing smaller or larger circles.
• Developed a simple artificial neural network model to simulate pigeon behavior.
• Analyzed the model's performance on novel stimulus pairs.

Main Results:

• The neural network model accurately reproduced pigeon behavior, including systematic failures.
• The model demonstrated generalization from trained to untrained stimulus pairs.
• The model's performance aligned with findings from Lazareva et al. (2014).

Conclusions:

• A simple artificial neural network can effectively model relational responding in pigeons.
• The model provides a computational account that integrates stimulus-feature and stimulus-relationship learning.
• The findings help reconcile Köhler's relational rules theory with Spence's stimulus generalization theory.