Generative Inferences Based on Learned Relations.

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Machine Learning

Background:

• Relational representations are key for generative inferences, allowing new conclusions from partial data.
• A challenge is explaining how generative inference arises from learning non-relational inputs.

Purpose of the Study:

• To demonstrate that a bottom-up relation learning model can achieve generative inference capabilities.
• To extend a model initially for comparative relation discrimination to perform generative tasks.

Main Methods:

• Utilized a bottom-up model for relation learning.
• Extended the model to handle generative inference tasks beyond simple discrimination.
• Evaluated the model's ability to make transitive inferences and account for human responses.

Main Results:

• The model successfully extended its relational learning to make generative inferences.
• The model demonstrated quasi-deductive transitive inference capabilities (e.g., A>B, B>C implies A>C).
• The model qualitatively matched human responses to generative questions about relative properties.

Conclusions:

• Bottom-up learning mechanisms in relational models can support generative inferences.
• This provides a computational basis for understanding how generative reasoning emerges from learning.
• The findings have implications for developing more sophisticated AI systems capable of relational reasoning.