A theory of relation learning and cross-domain generalization

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Developmental Psychology

Background:

• Human generalization of knowledge across diverse domains is a fundamental cognitive ability.
• Existing models often focus on statistical learning, potentially overlooking the role of structured representations.

Observation:

• A novel computational model, extending LISA and DORA, learns structured relational representations from nonrelational inputs without supervision.
• The model integrates reinforcement learning for domain-specific knowledge acquisition and analogical inference for zero-shot generalization.

Findings:

• The model successfully learns relational structures from visual stimuli and generalizes across video games (Breakout, Pong) and psychological tasks.
• Its learning trajectory closely mirrors children's developmental progression in reasoning and analogy-making.
• This demonstrates the power of abstract, relational representations for flexible cross-domain generalization.

Implications:

• This research provides a unified computational framework for understanding human generalization and analogy.
• It offers insights into the development of reasoning in children.
• The findings suggest that focusing on underlying relational structures, rather than surface statistics, is key for artificial intelligence and cognitive modeling.