An Active Inference Approach to Modeling Structure Learning: Concept Learning as an Example Case.

Area of Science:

• Computational Neuroscience
• Cognitive Science
• Machine Learning

Background:

• Structure learning, crucial for concept formation, involves both expanding internal models and reducing complexity.
• Existing models often struggle with generalization, data efficiency, and biological plausibility.

Purpose of the Study:

• To introduce a novel active inference framework for modeling structure learning, specifically state-space expansion and reduction.
• To demonstrate the potential of active inference in addressing limitations of current concept learning models.

Main Methods:

• Developed a generative model within the active inference framework equipped with adaptable 'slots' for new concept representation.
• Integrated a Bayesian model reduction process for simplifying models and increasing evidence.
• Utilized simulations to test the model's learning and generalization capabilities.

Main Results:

• Simulations showed the model could learn new concepts and refine existing ones with few observations.
• The approach demonstrated one-shot generalization to novel stimuli.
• Predicted neural underpinnings for these structure learning processes were simulated.

Conclusions:

• Active inference offers a promising framework for neurocomputational models of structure learning.
• The proposed approach provides a template for future research in real-world structure learning problems.
• This method enhances the development of biologically plausible and data-efficient learning algorithms.