Cross-dataset EEG decoding typically treats differences in montage, reference, and active sensors as a form of representation shift. A new arXiv paper argues that these changes also alter the observation operator—the mapping from brain activity to recorded signals—and should be handled accordingly. The authors propose GRAFT-EEG, a gauge-aware physical field transport approach that explicitly accounts for these physical changes during transfer.

The method is developed for target-statistics-blind transfer, meaning the model is fitted without access to any statistics from the target dataset. This constraint reflects real-world scenarios where target data may be unavailable or unlabeled at training time. By modeling the sensor configuration as part of the physical field rather than as a nuisance variable, GRAFT-EEG aims to improve decoding robustness across datasets.

The work is positioned as a step beyond conventional domain-adaptation techniques, which often ignore the physical basis of EEG measurement. While the abstract does not report quantitative results, the conceptual contribution is clear: treating sensor geometry and reference as part of the observation process, not just as distribution shift, may lead to more principled cross-dataset transfer for brain-computer interfaces.