Modeling Semantic Encoding in a Common Neural Representational Space.

Area of Science:

• Neuroscience
• Cognitive Science
• Machine Learning

Background:

• Individual semantic encoding models limit generalizability.
• Developing cross-subject models is crucial for broader applications.

Purpose of the Study:

• To develop and validate a novel method for estimating generalizable semantic encoding models across individuals.
• To compare the efficacy of anatomical normalization versus hyperalignment for between-subject modeling.

Main Methods:

• Functional MRI (fMRI) data acquired during naturalistic movie viewing.
• Word embeddings representing semantic content (agent, action, object, scene) were used.
• Conventional within-subject models, between-subject models with anatomical normalization, and between-subject models with hyperalignment were constructed.

Main Results:

• Anatomical normalization reduced spatial specificity in between-subject models.
• Hyperalignment preserved and recovered spatial specificity of semantic tuning.
• Hyperalignment-based models outperformed within-subject models in predictive performance.

Conclusions:

• Hyperalignment enables robust, generalizable semantic encoding models across individuals.
• This method enhances the prediction of brain responses to semantic information.
• Leveraging group data via hyperalignment improves out-of-sample prediction accuracy.