Learning abstract visual concepts via probabilistic program induction in a Language of Thought.

Area of Science:

• Cognitive Science
• Computational Linguistics
• Artificial Intelligence

Background:

• Abstract concept learning is crucial for human cognition.
• Variable binding is hypothesized to be key for abstract learning.
• Computational models of variable binding are underdeveloped.

Purpose of the Study:

• Formalize the Hierarchical Language of Thought (HLOT) model for rule learning.
• Investigate the computational aspects of variable binding in concept learning.
• Evaluate HLOT's ability to model human generalization patterns.

Main Methods:

• Developed the HLOT model using Bayesian inference and stochastic programs for variable binding.
• Designed an experiment where human subjects judged concept membership based on training items.
• Compared HLOT's predictions against variants of the Generalized Context Model (GCM).

Main Results:

• The HLOT model closely matched human generalization patterns.
• HLOT significantly outperformed GCM variants based on string and visual similarity.
• Evidence suggests variable binding occurs automatically, without adding cognitive complexity.

Conclusions:

• A hybrid approach combining symbolic variables with Bayesian inference and stochastic programs offers a novel perspective on generalization.
• The HLOT model provides a computationally grounded explanation for abstract concept learning.
• Variable binding is a fundamental mechanism in human generalization.