SPERA
An EEG foundation model incorporating spherical priors, based on a Joint Embedding Predictive Architecture (JEPA) that performs prediction in latent space. It achieves cross-device and cross-montage EEG representation learning through Legendre polynomial spatial priors and spatiotemporal decomposed attention mechanisms.
Timeline (last 90 days)
SPERA is an EEG foundation model that introduces a spherical prior and uses a Joint Embedding Predictive Architecture (JEPA) to perform predictions in latent space rather than reconstructing the observed signal.
The full name of SPERA is Spherical Prior EEG Representation Architecture.
SPERA introduces a spatial prior based on Legendre polynomials and integrates it into the attention mechanism to encode the geometric relationships among different scalp electrode arrangements.
SPERA designs factorized temporal and spatial attention, with full attention modules periodically interleaved to reduce computational complexity.
SPERA employs a relational spectral regularizer that aligns the similarity structure in latent space with spectral views to preserve sensitivity to frequency features.
SPERA is trained on approximately 80,000 hours of EEG data, covering 29,048 subjects across 106 datasets.
SPERA achieves the highest average balanced accuracy across nine downstream tasks spanning three application domains: clinical, cognitive, and brain-computer interface.
Under the linear probing setting, SPERA achieves excellent performance by training only a small number of downstream parameters.
SPERA has the potential to become a general-purpose pre-trained backbone for EEG analysis, provided its cross-device and cross-task generalization ability is validated in broader independent evaluations.
This article is compiled based on information from an arXiv preprint; the relevant conclusions are pending peer review and further replication validation.
All Facts (10)
This article is compiled based on information from an arXiv preprint; the relevant conclusions are pending peer review and further replication validation.
50%UnverifiedThe full name of SPERA is Spherical Prior EEG Representation Architecture.
50%UnverifiedSPERA introduces a spatial prior based on Legendre polynomials and integrates it into the attention mechanism to encode the geometric relationships among different scalp electrode arrangements.
50%UnverifiedSPERA designs factorized temporal and spatial attention, with full attention modules periodically interleaved to reduce computational complexity.
50%UnverifiedSPERA employs a relational spectral regularizer that aligns the similarity structure in latent space with spectral views to preserve sensitivity to frequency features.
50%UnverifiedSPERA is trained on approximately 80,000 hours of EEG data, covering 29,048 subjects across 106 datasets.
50%UnverifiedSPERA achieves the highest average balanced accuracy across nine downstream tasks spanning three application domains: clinical, cognitive, and brain-computer interface.
50%UnverifiedUnder the linear probing setting, SPERA achieves excellent performance by training only a small number of downstream parameters.
50%UnverifiedSPERA is an EEG foundation model that introduces a spherical prior and uses a Joint Embedding Predictive Architecture (JEPA) to perform predictions in latent space rather than reconstructing the observed signal.
50%UnverifiedSPERA has the potential to become a general-purpose pre-trained backbone for EEG analysis, provided its cross-device and cross-task generalization ability is validated in broader independent evaluations.
50%