Unsupervised method for representation transfer from one brain to another

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Neural information representation varies significantly across individuals.
• Decoding brain activity often requires personalized models, limiting data reusability.
• Brain-to-brain communication and cross-subject model application necessitate effective data translation.

Purpose of the Study:

• To develop an unsupervised method for brain representation transfer between individuals.
• To enable the transformation of neural data representations without corresponding label information.
• To demonstrate the applicability of the method for re-purposing existing brain decoding models.

Main Methods:

• Developed an algorithm for brain representation transfer using non-linear dimensional reduction via encoders.
• Utilized rotational and reflectional transformations on low-dimensional hyperspheres to capture common similarity structures.
• Validated the method using artificial neural network data and functional magnetic resonance imaging (fMRI) data from human participants.

Main Results:

• Successfully performed unsupervised brain representation transfer, achieving transformations comparable to supervised methods in some cases.
• Demonstrated the ability to reconstruct images from individual brain data without personalized decoders.
• Validated the method's effectiveness on both simulated neural activity and real fMRI data.

Conclusions:

• The proposed unsupervised transfer method is effective for re-applying participant-specific models to decode neural information from other individuals.
• This methodology serves as a proof of concept for exchanging latent neural information properties across individuals.
• Findings support the potential for broader application in personalized neuroscience and brain-computer interfaces.