Cortical knowledge structures guide word concept learning.

Area of Science:

• Cognitive Neuroscience
• Computational Neuroscience
• Neuroimaging

Background:

• Human word concept learning relies on prior knowledge for generalization from limited examples.
• The neural mechanisms underlying this process, particularly the role of prior knowledge, are not fully understood.

Purpose of the Study:

• To develop and test a Neural Bayesian Model (NBM) that integrates neural representational priors to explain word concept learning.
• To investigate the neural basis of concept learning by contrasting prior-based cortical mechanisms with exemplar-associative learning in the hippocampus.

Main Methods:

• Functional magnetic resonance imaging (fMRI) was used to map neural activity.
• A neural prior space was constructed from fMRI data of familiar objects.
• The NBM, incorporating these neural priors, was used to predict neural representations and behavioral generalization during novel word concept learning.

Main Results:

• The NBM, utilizing ventral occipitotemporal cortex priors, accurately predicted neural representations and behavioral generalization.
• The model outperformed control models that lacked neural priors.
• Hippocampal activity supported learning for novel shapes via prior-free associative mechanisms, distinct from the NBM's predictions.
• Large language models demonstrated less alignment with human generalization patterns compared to the NBM.

Conclusions:

• Cortical inference, guided by neural priors, underlies Bayesian concept acquisition.
• Hippocampal function supports exemplar-associative learning, separate from prior-based cortical mechanisms.
• This study clarifies the interaction between semantic memory (cortical priors) and episodic memory (hippocampal function) in concept learning.