Personalized visual encoding model construction with small data.

Area of Science:

• Neuroscience
• Machine Learning
• Brain Imaging

Background:

• Understanding individual differences in brain responses to stimuli is crucial for linking neural activity to behavior and pathology.
• Encoding models predict brain responses but typically require extensive functional magnetic resonance imaging (fMRI) data for accuracy.
• Existing methods struggle with data limitations for novel individuals.

Purpose of the Study:

• To develop an efficient ensemble approach for creating accurate personalized brain encoding models with limited fMRI data.
• To capture inter-individual variability in neural responses to stimuli.
• To enable personalized synthetic image generation for new subjects.

Main Methods:

• An ensemble method was proposed, modeling each subject's response as a linear combination of others' predicted responses.
• Models were trained and validated using hundreds of image-response pairs, comparing performance against models trained on 20,000 pairs.
• The approach was tested for robustness against domain shift (different scanner/setup) and applied to face-selective areas using the NeuroGen framework.

Main Results:

• Ensemble encoding models achieved accuracy comparable to models trained on significantly more data (20,000 pairs).
• The ensemble models successfully preserved inter-individual differences in the image-response relationship.
• The approach demonstrated robustness to domain shift and identified neural differences in face processing areas.

Conclusions:

• The proposed ensemble method efficiently creates accurate, personalized encoding models from limited fMRI data.
• This approach leverages existing densely-sampled datasets to overcome data scarcity for new individuals.
• The findings support the potential for personalized neuroscience and optimized synthetic stimuli generation.