USING CEBRA TO EXTRACT MEANINGFUL LATENT DYNAMICS FROM EEG RECORDINGS

Neuroscience is increasingly shifting from localized neural activity toward large-scale population dynamics. Although population activity is inherently high-dimensional, it is constrained by network connectivity to lower-dimensional manifolds. This neural manifold framework has enabled major advances in linking neural dynamics to behavioral outputs. Dimensionality reduction methods have successfully revealed such latent dynamics, but primarily from high signal-to-noise invasive recordings. Extending these approaches to non-invasive EEG, which exhibits substantial noise and inter-participant variability, remains challenging yet is critical for translational applications. We address this using CEBRA (Schneider et al., 2023), a self-supervised contrastive framework that jointly embeds neural and behavioral data, combined with a robust EEG encoder (Defossez et al., 2023) featuring spatial attention, participant-specific convolutions, and convolutional blocks. We evaluated the model on the Natural Objects Dataset–EEG (Zhang et al., 2025), where participants categorized ImageNet images as animate or inanimate (62-channel EEG, 250 Hz). Models trained on stimulus-period EEG using trial labels as auxiliary variables produced smooth, interpretable trajectories that diverged maximally around 250 ms post-stimulus, consistent with visual processing timescales, and reconverged by stimulus offset. Test trajectories closely matched training dynamics, indicating good generalization and potential applicability to BCI settings. Single-timepoint latent decoding achieved peak accuracy of 68% at ~370 ms (62–74% across subjects). Control models with shuffled labels or time failed to learn meaningful structure in latent trajectories. These findings demonstrate that CEBRA with robust encoding can extract behaviourally relevant neural dynamics from non-invasive EEG, similar to those revealed by invasive methods, bridging a critical translational gap.