Aligning Human and Mouse EEG representations of Sleep Stages with a Quadruplet Split Latent Permutation Auto-Encoder

Javier García Ciudad, Anders Vestergaard Nørskov, Alexander Neergaard Zahid, Birgitte Rahbek Kornum, Morten Mørup

Track: tiny / short paper (up to 5 pages)

Domain: machine learning

Abstract: Rodents are the most widely used experimental animals in biomedical research, but human neurological disorders involve behaviors and mental states that are challenging to study in rodents. Techniques such as electroencephalography (EEG) monitor brain activity objectively, yet how brain states manifest across species remains unclear, limiting translatability. We use a Quadruplet Split Latent Permutation Autoencoder (QSLP-AE) to map human and mouse sleep EEG into a shared 2-d latent space. The QSLP-AE exchanges latent representations during reconstruction to produce aligned representations of mouse and human sleep stages, as a proxy for cross-species representations of mental states. Notably, QSLP-AE matches the performance of conventional contrastive learning using only the autoencoder reconstruction loss. Exploiting the 2-d space, we visualize and quantify the correspondence between species from the model perspective. These results demonstrate the potential of QSLP-AE to align neural representations and bridge the translational gap.

Presenter: ~Javier_García_Ciudad1

Submission Number: 76