Decomposing dynamical subprocesses for compositional generalization.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Human Learning

Background:

• Human cognition excels at adapting to novel situations through abstraction and generalization.
• Limited understanding exists on how humans adapt to complex, multi-subprocess environments.

Purpose of the Study:

• To investigate if humans compositionally generalize knowledge of subprocess dynamics to solve new problems.
• To propose and test a computational mechanism for compositional generalization in human learning.

Main Methods:

• A novel sequence learning task involving compound images from graph product spaces was designed.
• Two participant groups learned distinct subprocesses, then transferred knowledge to a novel compound task.
• Computational modeling was used to compare different predictive representation theories.

Main Results:

• Participants demonstrated enhanced accuracy in predicting subprocess dynamics experienced during prior learning in a new task environment.
• Computational models without compositional transfer failed to explain the observed behavior.
• Behavior aligned with a model positing mapping of task states between learning phases.

Conclusions:

• Humans extract and generalize subprocesses compositionally, reusing prior knowledge for new experiences.
• This supports a mechanistic understanding of how cognitive flexibility enables adaptation to complex environments.
• Compositional generalization is a key mechanism for abstracting and reusing knowledge.