Compositional diversity in visual concept learning.

Area of Science:

• Cognitive Science
• Computer Vision
• Artificial Intelligence

Background:

• Humans efficiently learn new concepts by combining familiar parts (compositionality).
• Computer vision models often require extensive data and exhibit limited flexibility in generalizing concepts compared to humans.
• Understanding compositional generalization is key to developing more human-like AI.

Purpose of the Study:

• To investigate human abilities in visual composition for classifying and generating novel objects.
• To develop computational models that can replicate and explain human compositional generalization.
• To compare the performance of Bayesian program induction and neuro-symbolic program induction models against human behavior.

Main Methods:

• Experimental study of human classification and generation of "alien figures" with relational structure.
• Development of a Bayesian program induction model to infer generative programs for visual figures.
• Development of a neuro-symbolic program induction model incorporating neural network modules for residual structure.

Main Results:

• Both humans and the Bayesian program induction model demonstrated compositional generalization in few-shot classification tasks.
• The model provided a strong account of human data, revealing assumptions about invariant factors like rotation and part attachment.
• In few-shot generation, both humans and models created novel examples, though humans exhibited additional behaviors like set completion and part reconfiguration.

Conclusions:

• Humans and computational models can exhibit compositional behavior in visual classification and generation.
• Neuro-symbolic program induction models can capture additional human behavioral patterns beyond traditional Bayesian models.
• Findings advance the understanding of human concept learning and inform the development of more flexible AI systems.