Choose your track.

EEG epoch · 63 ch THINGS-EEG + Alljoined

rank

Held-out gallery

DINOv2 embedding space

Train and test images do not overlap Top-5 acc ↑

Visual decoding is most useful when it recognizes novel content rather than a fixed catalogue learned during training. Given a single EEG epoch recorded during natural-image viewing, identify the viewed image by ranking held-out candidates in a frozen DINOv2 embedding space. Training and test images do not overlap, so the cross-stimulus shift probes whether neural representations transfer beyond memorized concepts. Research-grade and consumer-grade datasets add source and hardware variability while keeping the task anchored to time-locked visual responses. Metric: Top-5 retrieval accuracy against the held-out candidate set, with higher scores indicating better performance.

A practical BCI must remain reliable as the same user returns on different days without requiring a new calibration each time. Decode one of three cued mental commands: kinesthetic motor imagery, mental calculation, or word association. The longitudinal corpus follows participants across six sessions and varied Graz and BrainHero contexts. Models train on earlier labeled sessions and predict later sessions. This within-user, cross-session design isolates drift from time of day, electrode replacement, physiology, and mental strategy while preserving subject-specific calibration. Metric: Balanced accuracy averaged across subject, session, and context cells, with higher scores indicating better performance.

Reliable sleep-onset estimates from lightweight home headbands could make longitudinal sleep monitoring less dependent on full laboratory polysomnography. Predict seconds remaining until the first stable N2 epoch from continuous four-channel wearable EEG collected at home. Models are evaluated on unseen sleepers, with night-to-night variation, motion artifacts, impedance changes, and channel dropout. The cross-user setting asks whether sleep-transition dynamics learned from other participants transfer to everyday recordings. It targets causal onset estimation rather than full hypnogram reconstruction, which is poorly supported by sparse wearable montages. Metric: Binned mean absolute error (bMAE) in seconds, averaged equally across four time-to-onset ranges so short and long latencies count equally, with lower scores indicating better performance.

Translating wrist muscle activity into hand motion is a core challenge for wearable control systems that must work beyond a calibrated laboratory setup. Predict trajectories of 20 hand-joint angles from 16-channel wrist sEMG recorded during everyday movement. Evaluation holds out users, movement stages, and user-stage combinations, testing robustness to anatomy, wristband placement, and kinematic context. This asks whether a wearable interface intended for daily, out-of-the-lab use can transfer to a new user without person-specific retraining. Metric: Mean absolute angular error across the predicted joint trajectories, reported in degrees, with lower scores indicating better performance.