Learning brain dynamics for decoding and predicting individual differences.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Cognitive, emotional, and motor functions rely on complex interactions within brain signals.
• Understanding these multivariate brain dynamics is crucial for decoding neural activity.

Purpose of the Study:

• To propose and validate a recurrent neural network architecture for uncovering distributed spatiotemporal signatures in brain dynamics.
• To decode brain activity during naturalistic conditions like movie watching and experimental paradigms.

Main Methods:

• Utilized functional Magnetic Resonance Imaging (fMRI) data from human participants.
• Developed a recurrent neural network model to learn spatiotemporal patterns from fMRI signals.
• Employed dimensionality reduction, saliency maps, and lesion analysis for interpretation.

Main Results:

• Achieved high accuracy (∼90%) in 15-way movie-clip classification at the brain region level.
• Successfully performed binary classification of experimental conditions (∼60%) at the voxel level.
• Modeled individual differences in fluid intelligence and verbal IQ, comparable to existing methods.

Conclusions:

• The proposed framework effectively decodes dynamic, spatially distributed brain representations.
• It offers a novel method for visualizing and analyzing neural dynamics during naturalistic tasks.
• This approach advances the understanding of how brain activity underlies complex cognitive functions.