Learning the generative principles of a symbol system from limited examples

Area of Science:

• Cognitive Science
• Developmental Psychology
• Artificial Intelligence

Background:

• Human learning mechanisms are debated, particularly if simple associative processes explain complex learning.
• Generative learning, or far generalization, is key to understanding how learners create new instances beyond experienced data.

Purpose of the Study:

• To investigate if imperfect, correlated data can support generative learning in children and AI.
• To explore the mechanisms behind far generalization in learning number names.

Main Methods:

• Two experimental studies with 148 preschool children learning number name-to-written form mappings.
• Computational modeling using a deep learning neural network trained on similar data.

Main Results:

• Both children and the neural network demonstrated systematic, far generalizations from imperfect data.
• Generative learning principles were evident despite limited examples and exceptions in training data.

Conclusions:

• Complex generalizations can emerge from imperfect, correlated data, challenging purely associative learning theories.
• Findings have implications for understanding human cognition, child development, education, and advancing artificial intelligence.