Multi-subject brain decoding with multi-task feature selection.

Area of Science:

• Neuroscience
• Machine Learning
• Brain Imaging

Background:

• Multi-subject brain decoding is crucial in neuroscience but challenging due to inter-subject variability in brain activation patterns.
• Existing methods struggle to effectively decode brain activity when using functional magnetic resonance imaging (fMRI) data pooled from multiple subjects.

Purpose of the Study:

• To develop a robust feature extraction method for effective multi-subject brain decoding using fMRI data.
• To address the challenge of individual brain variability in decoding models.
• To improve the accuracy of predicting visual stimuli from brain activity.

Main Methods:

• A hierarchical model was proposed to extract robust features for decoding.
• A novel multi-task feature selection method was introduced, treating feature selection for each subject as a separate task.
• This method leveraged complementary information across subjects and local correlations within brain areas.
• A linear Support Vector Machine (SVM) classifier was trained on pooled fMRI data for image prediction.

Main Results:

• The proposed hierarchical model and multi-task feature selection method demonstrated effectiveness in enhancing brain decoding.
• The approach successfully extracted robust features despite variations in brain activation patterns across subjects.
• The trained SVM classifier accurately predicted 2-D and 3-D stimuli-related images.

Conclusions:

• The developed hierarchical model offers a promising solution for effective multi-subject brain decoding.
• The multi-task feature selection approach successfully accounts for inter-subject variability and intra-subject correlations.
• This method advances the capability of predicting visual stimuli from fMRI data across diverse individuals.